CN110351318A - Using the method, terminal and computer storage medium of recommendation - Google Patents
Using the method, terminal and computer storage medium of recommendation Download PDFInfo
- Publication number
- CN110351318A CN110351318A CN201810301039.5A CN201810301039A CN110351318A CN 110351318 A CN110351318 A CN 110351318A CN 201810301039 A CN201810301039 A CN 201810301039A CN 110351318 A CN110351318 A CN 110351318A
- Authority
- CN
- China
- Prior art keywords
- application
- user
- feature
- sample
- recommendation
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/147—Network analysis or design for predicting network behaviour
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/50—Network services
- H04L67/535—Tracking the activity of the user
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/50—Network services
- H04L67/55—Push-based network services
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Computer Hardware Design (AREA)
- General Engineering & Computer Science (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
This application discloses method, terminal and computer storage mediums that a kind of application is recommended, include the first attributive character and the second attributive character in above-mentioned application feature this method comprises: obtaining the application feature of application to be recommended;It is determined according to the first attributive character of application to be recommended and applies recommendation network model, and by learning using recommendation network model to the second attributive character of application to be recommended, determine the corresponding application recommendation of application to be recommended or user's recommendation, wherein, it is obtained using recommendation network model by the sample application feature training of the first user-association, or is obtained by the sample of users feature training of the first association;According to above-mentioned using the determining application priority for recommending application to be recommended to the first user of recommendation, or the User Priority for recommending the first application is determined according to above-mentioned user's recommendation.Using the embodiment of the present application, the feasibility recommended using orientation can be improved, improve the accuracy rate that application is recommended.
Description
Technical field
This application involves method, a kind of terminal and a kind of meters that Internet technical field more particularly to a kind of application are recommended
Calculation machine storage medium.
Background technique
Function with terminals such as smart phone, tablet computers is become stronger day by day, and is applied to smart phone, tablet computer etc. eventually
Also become increasingly abundant application program (or referred to as applying) on end multiplicity therewith.It is brought using the diversification of type to terminal user
Rich and varied user experience, while also selection problem is added to terminal user.For convenience rapidly from diversified
The application for meeting end-user demands or hobby is quickly elected in, the recommendation service of types of applications is come into being.
In the prior art, the application that user had downloaded is generally based on using recommendation carry out the higher application of similarity
Recommend, demand either based on the user group for having similar tastes and interests or possessing common experience or likes to user and recommend to feel
The application of interest.However, the application either downloaded based on user, or demand or hobby based on user group
The recommendation for carrying out other application can make the application similarity elected higher, and the probability of occurrence of redundancy application is high, user
The application being not concerned with before this it is high by shielding probability it is poor for applicability so that the validity that application is recommended is low.
Summary of the invention
The embodiment of the present application provides a kind of method, terminal and computer storage medium that application is recommended, and can enhance to apply pushing away
The user-association recommended improves the feasibility that application orientation is recommended, improves the accuracy rate that application is recommended, and applicability is stronger.
In a first aspect, this application provides a kind of methods that application is recommended, this method comprises:
Obtain the application feature of application to be recommended, wherein include the first attribute in the application feature of above-mentioned application to be recommended
Feature and the second attributive character, above-mentioned first attributive character is the first user property feature, above-mentioned second attributive character is second
Application attribute feature or above-mentioned first attributive character are the first application attribute feature, above-mentioned second attributive character is the second use
Family attributive character;
It is determined according to the first attributive character of above-mentioned application to be recommended and applies recommendation network model, and pushed away by above-mentioned application
It recommends network model to learn the second attributive character of above-mentioned application to be recommended, determines the corresponding application of above-mentioned application to be recommended
Recommendation or the corresponding user's recommendation of above-mentioned application to be recommended, wherein above-mentioned application recommendation network model is by the first user
Associated sample application feature training obtains or the above-mentioned sample of users spy using recommendation network model by the first association
Sign training obtains;
Recommend above-mentioned application to be recommended according to determining using recommendation to above-mentioned first user for above-mentioned application to be recommended
Using priority, or is determined according to user's recommendation of above-mentioned application to be recommended and recommend the user of above-mentioned first application preferential
Grade.
In one possible implementation, the above method further include:
When the application priority of above-mentioned application to be recommended be more than or equal to it is default using priority threshold value when, Xiang Shangshu the
One user recommends above-mentioned application to be recommended;Or
When the User Priority of above-mentioned application to be recommended is more than or equal to pre-set user priority threshold value, determine to the
Two users recommend above-mentioned first application.
In one possible implementation, above-mentioned first attributive character is the first user property feature, above-mentioned second category
Property feature be the second application attribute feature;
It is above-mentioned to include: using recommendation network model according to the determination of the first attributive character of above-mentioned application to be recommended
Each application recommendation network mould that will include in above-mentioned first user property feature and application recommendation network model set
The associated user property feature of type is matched, and determines that above-mentioned first user belongs to from above-mentioned application recommendation network model set
Property corresponding above-mentioned first user of feature associated by apply recommendation network model;
Wherein, above-mentioned using the user in recommendation network model set further including other users except above-mentioned first user
Other application recommendation network model associated by attributive character, above-mentioned other application recommendation network model are closed by above-mentioned other users
The sample application feature training of connection obtains.
In one possible implementation, before the application feature of above-mentioned acquisition application to be recommended, the above method is also wrapped
It includes:
Obtain the sample data for recommending at least two samples application of training for application, wherein any sample application
It include above-mentioned first user attribute data and sample application attribute data in sample data;
At least one sample application feature pair is constructed according to the sample data that above-mentioned at least two sample is applied, wherein one
A sample application feature centering includes a positive sample feature and a negative sample feature, wherein is wrapped in above-mentioned positive sample feature
Above-mentioned first user property feature and positive sample application attribute feature are included, includes that above-mentioned first user belongs in above-mentioned negative sample feature
Property feature and negative sample application attribute feature;
Recommendation network model is applied to building according at least one above-mentioned sample application feature.
It in one possible implementation, include active degree in the sample application attribute data of above-mentioned various kinds this application
Indicate information;
The above-mentioned sample data according to the application of above-mentioned at least two sample constructs at least one sample application feature to including:
The application of above-mentioned at least two sample is matched two-by-two, obtains the application pair of at least one sample;
At least one above-mentioned any sample of sample application centering is applied, i is performed the following operations to obtain at least one sample
Using feature pair:
According to above-mentioned sample apply to the active degree instruction information of two samples of i application determine positive sample apply and
Negative sample application, wherein the active degree of above-mentioned positive sample application is higher than the active degree of above-mentioned negative sample application;
Positive sample is constructed according to the sample application attribute data that above-mentioned first user attribute data and above-mentioned positive sample are applied
Feature i1, and negative sample is constructed according to the sample application attribute data that above-mentioned first user attribute data and above-mentioned negative sample are applied
Feature i0, obtain above-mentioned sample and apply to the corresponding sample application feature of i to i10;
Wherein, above-mentioned sample application feature is to i10In include above-mentioned positive sample feature i1With above-mentioned negative sample feature i0。
In one possible implementation, above-mentioned that building application is recommended according at least one above-mentioned sample application feature
Network model includes:
The positive sample feature of at least one above-mentioned sample application feature centering various kinds this application feature pair and negative sample is special
Sign as using recommendation network model input, by it is above-mentioned using recommendation network model to above-mentioned various kinds this application feature pair
Positive sample feature and negative sample feature are learnt, any using the corresponding ability using recommendation of feature to obtain prediction;
Wherein, the corresponding application recommendation of any sample application feature centering positive sample feature is corresponding greater than negative sample feature
Application recommendation.
In one possible implementation, above by above-mentioned using recommendation network model to above-mentioned application to be recommended
Second attributive character is learnt, and determines that the corresponding application recommendation of above-mentioned application to be recommended includes:
Second application attribute feature of above-mentioned application to be recommended and above-mentioned first user property feature are inputted into above-mentioned application
Recommendation network model learns above-mentioned second application attribute feature by above-mentioned application recommendation network model, and export to
Corresponding above-mentioned first user of above-mentioned first user property feature recommends the application recommendation of above-mentioned application to be recommended.
In one possible implementation, any of the above-described sample is applied pushes away to positive sample application corresponding first is default in i
Recommend value, the corresponding second default recommendation of negative sample application;
By any sample application feature to i10Positive sample feature i1With negative sample feature i0Above-mentioned application is inputted to recommend
After network model, the above method further include:
Obtain the above-mentioned positive sample feature i exported using recommendation network model1Corresponding first applies recommendation and negative sample
Eigen i0Corresponding second applies recommendation;
According to the difference of above-mentioned first application recommendation and above-mentioned second application recommendation, in conjunction with the above-mentioned first default recommendation
The difference of value and the above-mentioned second default recommendation calculates recommendation loss;
Above-mentioned application recommendation network model is corrected according to the loss of above-mentioned recommendation, adjusts above-mentioned application recommendation network model pair
It is any using the corresponding precision of prediction using recommendation.
In one possible implementation, above-mentioned first attributive character is the first application attribute feature, above-mentioned second category
Property feature be second user attributive character;
It is above-mentioned to include: using recommendation network model according to the determination of the first attributive character of above-mentioned application to be recommended
Each application recommendation network mould that will include in above-mentioned first application attribute feature and application recommendation network model set
The associated application attribute feature of type is matched, and determines that above-mentioned first application belongs to from above-mentioned application recommendation network model set
Property corresponding above-mentioned first application of feature is associated applies recommendation network model;
Wherein, it is above-mentioned using further include in recommendation network model set it is above-mentioned first application except other application application
Other application recommendation network model associated by attributive character, above-mentioned other application recommendation network model are closed by above-mentioned other application
The sample of users feature training of connection obtains.
In one possible implementation, before the application feature of above-mentioned acquisition application to be recommended, the above method is also wrapped
It includes:
Obtain the user data for recommending at least two sample of users of training for application, wherein any sample of users
It include above-mentioned first application attribute data and sample of users attribute data in user data;
At least one sample of users feature pair is constructed according to the user data of above-mentioned at least two sample of users, according to above-mentioned
At least one sample of users feature applies recommendation network model to building, wherein a sample of users feature centering includes one
Positive sample feature and a negative sample feature, wherein including above-mentioned first application attribute feature and just in above-mentioned positive sample feature
Sample of users attributive character includes that above-mentioned first application attribute feature and negative sample user property are special in above-mentioned negative sample feature
Sign.
It in one possible implementation, include active degree in the sample of users attribute data of above-mentioned each sample of users
Indicate information;
The above-mentioned user data according to above-mentioned at least two sample of users constructs at least one sample of users feature to including:
Above-mentioned at least two sample of users is matched two-by-two, obtains at least one sample of users pair;
At least one above-mentioned any sample of users of sample of users centering performs the following operations i to obtain at least one sample
User characteristics pair:
According to above-mentioned sample of users to the active degrees of two sample of users of i instruction information determine positive sample user and
Negative sample user, wherein the active degree of above-mentioned positive sample user is higher than the active degree of above-mentioned negative sample user;
Positive sample is constructed according to the sample of users attribute data of above-mentioned first application attribute data and above-mentioned positive sample user
Feature i1, and negative sample is constructed according to the sample of users attribute data of above-mentioned first application attribute data and above-mentioned negative sample user
Feature i0, above-mentioned sample of users is obtained to the corresponding sample of users feature of i to i10;
Wherein, above-mentioned sample of users feature is to i10In include above-mentioned positive sample feature i1With above-mentioned negative sample feature i0。
In one possible implementation, above-mentioned that building application is recommended according at least one above-mentioned sample of users feature
Network model includes:
The positive sample feature of at least one above-mentioned each sample user characteristics pair of sample of users feature centering and negative sample is special
Sign as using recommendation network model input, by it is above-mentioned using recommendation network model to above-mentioned each sample user characteristics pair
Positive sample feature and negative sample feature are learnt, and predict any ability using the corresponding user's recommendation of feature to obtain;
Wherein, the corresponding user's recommendation of any sample of users feature centering positive sample feature is corresponding greater than negative sample feature
User's recommendation.
In one possible implementation, above by above-mentioned using recommendation network model to above-mentioned application to be recommended
Second attributive character is learnt, and determines that the corresponding user's recommendation of above-mentioned application to be recommended includes:
The second user attributive character of above-mentioned application to be recommended and above-mentioned first application attribute feature are inputted into above-mentioned application
Recommendation network model learns above-mentioned second user attributive character by above-mentioned application recommendation network model, and export to
The corresponding above-mentioned second user of above-mentioned second user attributive character recommends user's recommendation of above-mentioned first application.
In one possible implementation, any of the above-described sample of users corresponding to positive sample user in i first is default pushes away
Recommend value, the corresponding second default recommendation of negative sample user;
By any sample of users feature to i10Positive sample feature i1With negative sample feature i0Above-mentioned application is inputted to recommend
After network model, the above method further include:
Obtain the above-mentioned positive sample feature i exported using recommendation network model1Corresponding first user recommendation and negative sample
Eigen i0Corresponding second user recommendation;
According to the difference of above-mentioned first user recommendation and above-mentioned second user recommendation, in conjunction with the above-mentioned first default recommendation
The difference of value and the above-mentioned second default recommendation calculates recommendation loss;
Above-mentioned application recommendation network model is corrected according to the loss of above-mentioned recommendation, adjusts above-mentioned application recommendation network model pair
The precision of prediction of the corresponding user's recommendation of any user.
In one possible implementation, the user property feature of above-mentioned first user and/or above-mentioned second user by
Age of user, user's gender, user's educational background, region and user locating for user are using any user attribute data in account
It determines;
The sample application attribute feature of application attribute feature and/or the sample application of above-mentioned application to be recommended is marked by application
At least one of knowledge, application type, active degree instruction information, application resource type and user behavior data application attribute
Data determine.
Second aspect, this application provides a kind of terminal, which includes:
Feature acquiring unit, for obtaining the application feature of application to be recommended, wherein the application of above-mentioned application to be recommended is special
It include the first attributive character and the second attributive character in sign, above-mentioned first attributive character is the first user property feature, above-mentioned the
Two attributive character are the second application attribute feature or above-mentioned first attributive character is the first application attribute feature, above-mentioned second
Attributive character is second user attributive character;
First attribute of characteristic processing unit, the above-mentioned application to be recommended for being obtained according to features described above acquiring unit is special
Sign, which determines, applies recommendation network model, and by above-mentioned special using second attribute of the recommendation network model to above-mentioned application to be recommended
Sign is learnt, and determines that the corresponding application recommendation of above-mentioned application to be recommended or the corresponding user of above-mentioned application to be recommended are recommended
Value, wherein above-mentioned trained to be obtained or above-mentioned application using recommendation network model by sample application the feature of the first user-association
Recommendation network model is obtained by the sample of users feature training of the first association;
Recommend predicting unit, the above-mentioned application recommendation for determining according to features described above processing unit is determined to above-mentioned the
One user recommends the application priority of above-mentioned application to be recommended, or is pushed away according to the above-mentioned user that features described above processing unit determines
It recommends value and determines the User Priority for recommending above-mentioned first application.
In one possible implementation, above-mentioned recommendation predicting unit, is also used to:
When the application priority of above-mentioned application to be recommended be more than or equal to it is default using priority threshold value when, Xiang Shangshu the
One user recommends the application to be recommended;Or
When the User Priority of above-mentioned application to be recommended is more than or equal to pre-set user priority threshold value, determine to the
Two users recommend above-mentioned first application.
In one possible implementation, above-mentioned first attributive character is the first user property feature, above-mentioned second category
Property feature be the second application attribute feature;
Features described above processing unit is used for:
To include in the first user property feature that features described above acquiring unit obtains and application recommendation network model set
It is each matched using the user property feature of recommendation network model interaction, from above-mentioned application recommendation network model set really
It makes and applies recommendation network model associated by corresponding above-mentioned first user of above-mentioned first user property feature;
Wherein, above-mentioned using the user in recommendation network model set further including other users except above-mentioned first user
Other application recommendation network model associated by attributive character, above-mentioned other application recommendation network model are closed by above-mentioned other users
The sample application feature training of connection obtains.
In one possible implementation, features described above acquiring unit is also used to:
Obtain the sample data for recommending at least two samples application of training for application, wherein any sample application
It include above-mentioned first user attribute data and sample application attribute data in sample data;
At least one sample application feature pair is constructed according to the sample data that above-mentioned at least two sample is applied, wherein one
A sample application feature centering includes a positive sample feature and a negative sample feature, wherein is wrapped in above-mentioned positive sample feature
Above-mentioned first user property feature and positive sample application attribute feature are included, includes that above-mentioned first user belongs in above-mentioned negative sample feature
Property feature and negative sample application attribute feature;
Recommendation network model is applied to building according at least one above-mentioned sample application feature.
It in one possible implementation, include active degree in the sample application attribute data of above-mentioned various kinds this application
Indicate information;
Features described above acquiring unit is used for:
The application of above-mentioned at least two sample is matched two-by-two, obtains the application pair of at least one sample;
At least one above-mentioned any sample of sample application centering is applied, i is performed the following operations to obtain at least one sample
Using feature pair:
According to above-mentioned sample apply to the active degree instruction information of two samples of i application determine positive sample apply and
Negative sample application, wherein the active degree of above-mentioned positive sample application is higher than the active degree of above-mentioned negative sample application;
Positive sample is constructed according to the sample application attribute data that above-mentioned first user attribute data and above-mentioned positive sample are applied
Feature i1, and negative sample is constructed according to the sample application attribute data that above-mentioned first user attribute data and above-mentioned negative sample are applied
Feature i0, obtain above-mentioned sample and apply to the corresponding sample application feature of i to i10;
Wherein, above-mentioned sample application feature is to i10In include above-mentioned positive sample feature i1With above-mentioned negative sample feature i0。
In one possible implementation, features described above acquiring unit is used for:
The positive sample feature of at least one above-mentioned sample application feature centering various kinds this application feature pair and negative sample is special
Sign as using recommendation network model input, by it is above-mentioned using recommendation network model to above-mentioned various kinds this application feature pair
Positive sample feature and negative sample feature are learnt, any using the corresponding ability using recommendation of feature to obtain prediction;
Wherein, the corresponding application recommendation of any sample application feature centering positive sample feature is corresponding greater than negative sample feature
Application recommendation.
In one possible implementation, features described above processing unit is used for:
Second application attribute feature of above-mentioned application to be recommended and above-mentioned first user property feature are inputted into above-mentioned application
Recommendation network model learns above-mentioned second application attribute feature by above-mentioned application recommendation network model, and export to
Corresponding above-mentioned first user of above-mentioned first user property feature recommends the application recommendation of above-mentioned application to be recommended.
In one possible implementation, any of the above-described sample is applied pushes away to positive sample application corresponding first is default in i
Recommend value, the corresponding second default recommendation of negative sample application;
Features described above processing unit is also used to:
Obtain the above-mentioned positive sample feature i exported using recommendation network model1Corresponding first applies recommendation and negative sample
Eigen i0Corresponding second applies recommendation;
According to the difference of above-mentioned first application recommendation and above-mentioned second application recommendation, in conjunction with the above-mentioned first default recommendation
The difference of value and the above-mentioned second default recommendation calculates recommendation loss;
Above-mentioned application recommendation network model is corrected according to the loss of above-mentioned recommendation, adjusts above-mentioned application recommendation network model pair
It is any using the corresponding precision of prediction using recommendation.
In one possible implementation, above-mentioned first attributive character is the first application attribute feature, above-mentioned second category
Property feature be second user attributive character;
Features described above processing unit is used for:
Each application recommendation network mould that will include in above-mentioned first application attribute feature and application recommendation network model set
The associated application attribute feature of type is matched, and determines that above-mentioned first application belongs to from above-mentioned application recommendation network model set
Property corresponding above-mentioned first application of feature is associated applies recommendation network model;
Wherein, it is above-mentioned using further include in recommendation network model set it is above-mentioned first application except other application application
Other application recommendation network model associated by attributive character, above-mentioned other application recommendation network model are closed by above-mentioned other application
The sample of users feature training of connection obtains.
In one possible implementation, features described above acquiring unit is also used to:
Obtain the user data for recommending at least two sample of users of training for application, wherein any sample of users
It include above-mentioned first application attribute data and sample of users attribute data in user data;
At least one sample of users feature pair is constructed according to the user data of above-mentioned at least two sample of users, according to above-mentioned
At least one sample of users feature applies recommendation network model to building, wherein a sample of users feature centering includes one
Positive sample feature and a negative sample feature, wherein including above-mentioned first application attribute feature and just in above-mentioned positive sample feature
Sample of users attributive character includes that above-mentioned first application attribute feature and negative sample user property are special in above-mentioned negative sample feature
Sign.
It in one possible implementation, include active degree in the sample of users attribute data of above-mentioned each sample of users
Indicate information;
Features described above acquiring unit is used for:
Above-mentioned at least two sample of users is matched two-by-two, obtains at least one sample of users pair;
At least one above-mentioned any sample of users of sample of users centering performs the following operations i to obtain at least one sample
User characteristics pair:
According to above-mentioned sample of users to the active degrees of two sample of users of i instruction information determine positive sample user and
Negative sample user, wherein the active degree of above-mentioned positive sample user is higher than the active degree of above-mentioned negative sample user;
Positive sample is constructed according to the sample of users attribute data of above-mentioned first application attribute data and above-mentioned positive sample user
Feature i1, and negative sample is constructed according to the sample of users attribute data of above-mentioned first application attribute data and above-mentioned negative sample user
Feature i0, above-mentioned sample of users is obtained to the corresponding sample of users feature of i to i10;
Wherein, above-mentioned sample of users feature is to i10In include above-mentioned positive sample feature i1With above-mentioned negative sample feature i0。
In one possible implementation, features described above acquiring unit is used for:
The positive sample feature of at least one above-mentioned each sample user characteristics pair of sample of users feature centering and negative sample is special
Sign as using recommendation network model input, by it is above-mentioned using recommendation network model to above-mentioned each sample user characteristics pair
Positive sample feature and negative sample feature are learnt, and predict any ability using the corresponding user's recommendation of feature to obtain;
Wherein, the corresponding user's recommendation of any sample of users feature centering positive sample feature is corresponding greater than negative sample feature
User's recommendation.
In one possible implementation, features described above processing unit is used for:
The second user attributive character of above-mentioned application to be recommended and above-mentioned first application attribute feature are inputted into above-mentioned application
Recommendation network model learns above-mentioned second user attributive character by above-mentioned application recommendation network model, and export to
The corresponding above-mentioned second user of above-mentioned second user attributive character recommends user's recommendation of above-mentioned first application.
In one possible implementation, any of the above-described sample of users corresponding to positive sample user in i first is default pushes away
Recommend value, the corresponding second default recommendation of negative sample user;
Features described above processing unit is also used to:
Obtain the above-mentioned positive sample feature i exported using recommendation network model1Corresponding first user recommendation and negative sample
Eigen i0Corresponding second user recommendation;
According to the difference of above-mentioned first user recommendation and above-mentioned second user recommendation, in conjunction with the above-mentioned first default recommendation
The difference of value and the above-mentioned second default recommendation calculates recommendation loss;
Above-mentioned application recommendation network model is corrected according to the loss of above-mentioned recommendation, adjusts above-mentioned application recommendation network model pair
The precision of prediction of the corresponding user's recommendation of any user.
In one possible implementation, the user property feature of above-mentioned first user and/or second user is by user
Age, user's gender, user's educational background, region and user locating for user are true using any user attribute data in account
It is fixed;
The sample application attribute feature of application attribute feature and/or the sample application of above-mentioned application to be recommended is marked by application
At least one of knowledge, application type, active degree instruction information, application resource type and user behavior data application attribute
Data determine.
The third aspect, present invention also provides a kind of computer storage medium, above-mentioned computer storage medium is stored with more
Item instruction, when described instruction is run at the terminal, so that above-mentioned terminal executes above-mentioned first aspect and/or first aspect is any
The method that the possible implementation of kind provides.
Fourth aspect, present invention also provides a kind of terminal, including processor and memory, above-mentioned processor and memory
Be connected with each other, wherein above-mentioned memory for store support terminal execute above-mentioned first aspect and/or first aspect is any can
The computer program for the method that the implementation of energy provides, above-mentioned computer program includes program instruction, and above-mentioned processor is matched
It sets for calling above procedure to instruct, executes above-mentioned first aspect and/or any possible implementation of first aspect provides
Method.
The application attribute feature that the user property feature of first user and/or first are applied can be used for by the embodiment of the present application
Application feature and/or user in the building using feature of first application, by application recommendation network model to the first application
Feature is learnt to be recommended with pre- the first user of direction finding recommendation when the first application and/or says that first recommends the using orientation
The recommendation of two users, and then can realize and the orientation application of the first user is recommended and/or the directional user of the first application is pushed away
It recommends, it is easy to operate.The application attribute feature etc. that the embodiment of the present application applies the user property feature of the first user and/or first
Be dissolved into the recommendation process of the first application, enhance first application recommend with first apply and/or the first user be associated with it is close
Cutting property, and then the probability for recommending the game of non-user demand or hobby to the first user can be reduced, improve the standard that application is recommended
True rate, while the user's viscosity for enhancing terminal using the redundancy rate recommended can also be reduced.
Detailed description of the invention
It, below will be to required use in the embodiment of the present application description in order to illustrate more clearly of the technical solution of the application
Attached drawing be briefly described.
Fig. 1 is the application scenarios schematic diagram provided by the embodiments of the present application using recommended method;
Fig. 2 is the flow diagram provided by the embodiments of the present application using recommended method;
Fig. 3 is the another application schematic diagram of a scenario provided by the embodiments of the present application using recommended method;
Fig. 4 is another flow diagram provided by the embodiments of the present application using recommended method;
Fig. 5 is the building schematic diagram provided by the embodiments of the present application using recommendation network model;
Fig. 6 is another flow diagram provided by the embodiments of the present application using recommended method;
Fig. 7 is another flow diagram provided by the embodiments of the present application using recommended method;
Fig. 8 is a structural schematic diagram of terminal provided by the embodiments of the present application;
Fig. 9 is another structural schematic diagram of terminal provided by the embodiments of the present application.
Specific embodiment
Below in conjunction with the attached drawing provided in the embodiment of the present application, technical solution provided by the embodiments of the present application is carried out clear
Chu is fully described by.
The method that application provided by the embodiments of the present application is recommended (may be simply referred to as convenience of description using recommended method or side
Method) terminal that is applicable in includes but is not limited to smart phone, computer, tablet computer, personal digital assistant (personal
Digital assistant, PDA), mobile internet device (mobile Internet device, MID) and wearable set
It is standby etc..Optionally, terminal can also be corresponding for above-mentioned smart phone, computer, tablet computer, PDA, MID and wearable device
Server etc., can specifically be determined according to practical application scene, herein with no restrictions.For convenience of description, the embodiment of the present application is mentioned
The executing subject using recommended method supplied will be illustrated with terminal.Corresponding, application provided by the embodiments of the present application is recommended
Device (or referred to as apply recommendation apparatus) include but is not limited to smart phone, computer, tablet computer, PDA, MID and can
Wearable device etc..For convenience of description, application recommendation apparatus provided by the embodiments of the present application and/or terminal will with smart phone (or
Abbreviation mobile phone) for be illustrated.
The recommendation provided by the embodiments of the present application that a plurality of types of applications are applicable to using recommended method, including appoint
It anticipates the recommendation of a type of a plurality of applications or the recommendation of a plurality of types of a plurality of applications, herein with no restrictions.Wherein, on
State a plurality of types of applications (such as mobile phone application) including but not limited to: game class application, healthy class are applied, shopping class is applied,
Tool-class application, multimedia class application, social category application, the application of travelling class and educational application etc., herein with no restrictions.Its
In, it may include a plurality of applications in the application of the above-mentioned same type, herein with no restrictions.For example, above-mentioned game class apply including
But it is not limited to king's honor, QQ driving, happy fighting landlord and cruel race etc. everyday.Above-mentioned health class application includes but is not limited to strong
Body application, cuisines application and vital sign (such as blood pressure) record application etc..Above-mentioned shopping class application includes but is not limited to shadow
Application etc. is purchased depending on booking application, cuisines booking application and daily necessities.Above-mentioned tool-class application includes but is not limited to text
Part editor, mail, alarm clock, calendar, photograph album, setting and compass etc..Above-mentioned multimedia class application may include video player,
Music player, photography applications, the application of U.S. figure and audio typing application etc..Above-mentioned social application class includes but is not limited to wechat
And QQ etc..Above-mentioned travelling class is applied including railway 12306, oozes and go and take journey travelling etc..Above-mentioned educational application packet
It includes but is not limited to wechat is read and QQ is read etc..
Optionally, various types of applications in the example above can individualism, mutually nested can also exist, not do herein
Limitation.Wherein, above-mentioned individualism can be understood as being the application in application library (such as application shop etc.), by application library into
A plurality of types of applications that row classification obtains.For example, above-mentioned game class application can be the game in application library, such as application shop
In a plurality of game etc..The application that above-mentioned mutually nested presence can be understood as a certain seed type can be in some application
One of application function.For example, above-mentioned game class application can be with social category using mutually nested, so that game class, which is applied, becomes society
Hand over one of the application function of class application.In other words, for example, the wechat in the application of above-mentioned social category can be with the king in game class
A plurality of game such as honor are mutually nested, for example, this application function (abbreviation wechat trip of settable game in the application function of wechat
Play), and may include a plurality of game such as king's honor, happy fighting landlord in wechat game.
For convenience of description, it is provided by the embodiments of the present application using recommended method be applicable in application will be with game class application
It is illustrated for (abbreviation game), i.e., application recommended method provided by the embodiments of the present application concretely game recommdation method.
It is corresponding, the application scenarios provided by the embodiments of the present application that a variety of game recommdations are applicable to using recommended method, wherein can wrap
The recommendation for including any money game in application library, being equally applicable to any application (can be set as applying A for convenience of description, such as micro-
Letter) in any money game recommendation, herein with no restrictions.Below by by application scenarios be using the game recommdation in A for,
It is illustrated to provided by the embodiments of the present application using the applicable scene of recommended method in conjunction with Fig. 1.
It is the application scenarios schematic diagram provided by the embodiments of the present application using recommended method referring to Fig. 1, Fig. 1.In this Shen
Please be in embodiment, terminal (such as mobile phone) can monitor the user's operation state in its user interface (such as interface 1).It is false such as Fig. 1
If be mounted in terminal using A, using B, using terminal applies such as C.When terminal user (abbreviation user) clicks the user of terminal
When applying the icon of A on interface, the user interface (such as interface 2) that A is applied in terminal starting can trigger.At this point, terminal can monitor
User operation instruction onto its user interface, and can determine that user selects triggering according to the click location of user operation instruction
The application of starting is using A.At this point, terminal can start the user interface (such as interface 2) using A.As shown in Figure 1, it is assumed that
Using including on multiple changeable interfaces such as chat, contact person and extension function in the user interface of A, user can be clicked
Switch the function interface different using A using the different icons in the user interface of A.For example, it is assumed that user clicks extension function
The icon of energy, terminal detect the user interface that then can be switched when the user operation instruction in extension function corresponding icon using A
A variety of extension functions of A are applied in display, for example, application function 1, application function 2 ..., game etc..Further, when the user clicks
When operating area corresponding using game in the extension function of A, the user interface of terminal switch application A can trigger to game pair
The display interface (such as interface 3) answered.On interface 3, the temperature ranking list of a plurality of game can be shown, and then can recommend to user
Corresponding game.
A plurality of game built in application A are recommended as shown in Figure 1, a variety of ordering rules can be shown as on interface 3
Recommendation list, including heat trip list, good friend's heat plays list and new trip list etc..When terminal detects on any display area for recommending list
User operation instruction when, terminal switches to user operation instruction and corresponds to selected recommendation list, will recommend list on game push away
It recommends list output and shows user.For example, when the user clicks when " I " corresponding operating area, terminal then can switch to " I
" corresponding recommend list for the personal of user's recommended games.In the specific implementation, how terminal is directed to of user's recommended games
People recommends list to can be found in implementation described in following each embodiment.
Optionally, in the embodiment of the present application, game recommdation when can also according to each money game user group's quantity or
User pays close attention to user groups' parameters such as temperature and is ranked up a plurality of game, and then the forward a plurality of game recommdations that will sort are to use
Family.Such as it is shown in FIG. 1, a plurality of game are ranked up according to user group's parameter to obtain heat trip list, or close according to user
The a plurality of game newly released are ranked up to obtain new trip list by the user groups such as heat injection degree parameter, and then can be by forward 5 of sorting
User is more recommended in money game.Alternatively, the good friend according to user joins the good friends groups such as the concern temperature of each money game
A plurality of game are ranked up by number, obtain good friend's heat trip list, and by 5 sections of game of the umber one or more recommend user.Specifically
It can be determined according to practical application scene, herein with no restrictions.
However, a plurality of game are ranked up according to user groups' parameters such as the concern temperatures of user group to obtain recommendation column
Table and when by ranking forward game recommdation to user, both may include that user had paid attention to and playing in recommendation list
Game, may also comprise with user in other the higher game of the game similarity of object for appreciation, however these recommendations are frequently not
User is interested or the game that needs, so that the recommendation probability of redundancy game is high.In addition, for be not awfully hot door game,
Or the new trip (i.e. trendy game) not being concerned about by user also, be often difficult to appear in and recommend in ranking list, so be difficult to by
User's discovery, the exposure rate of game is low, and also reduces user and the diversity of game may be selected.
It is pushed away based on above-mentioned be ranked up a plurality of game according to user groups' parameters such as the concern temperatures of user group
The height of redundancy probability present in list is recommended, defects, the embodiment of the present application such as accuracy rate is low is recommended to provide a kind of achievable game
Directional user recommend method.Game recommdation method provided by the embodiments of the present application can utilize some user (such as first
User) user property feature and the first user played the game characteristics such as the game attributes feature of game carry out game recommdation
The building of model, and then the game attributes spy for the game played using the user property feature of the first user and the first user
Sign etc. is trained the game recommdation model so that game recommdation model can be for each in any a or a plurality of game
The game characteristic of money game determines the recommendation for recommending each money game to the first user.It is determined by the height difference of recommendation
When recommending each money game to target, the recommended priority of each money game.Wherein, the higher game of recommendation is corresponding recommends preferentially
Grade is higher.The game attributes for the game that the embodiment of the present application can play the user property feature of the first user and the first user
The features such as feature are associated with the recommendation of other any game foundation, and then can realize the orientation recommendation to the first user, operation letter
It is single, and then the probability for recommending the game of non-user demand or hobby to the first user can be reduced, improve the accurate of game recommdation
Rate, while the redundancy rate of game recommdation can also be reduced, the user's viscosity for enhancing terminal.
Optionally, game recommdation method provided by the embodiments of the present application some game can also be used (can for convenience of description
Referred to as the first game) game attributes feature and the first game the game characteristics such as the user property feature of user group
The building of game recommdation model is carried out, and then using the game attributes feature of the first game and the user group of the first game
User property feature etc. the game recommdation model is trained so that game recommdation model can for any a or
The application feature of each money game in a plurality of game, which determines, recommends the first game orientation in any one or multiple users respectively
The recommendation of user.Determine that each user is corresponding to push away when recommending the first game to each user by the height difference of recommendation
Recommend priority.Wherein, the corresponding recommended priority of the higher user of recommendation is higher.The embodiment of the present application can be by the first game
Game attributes feature and the user property feature of the first game (i.e. second user attributive character) etc. apply feature, with other
Association is established in the recommendation of one user, and then can realize the first game orientation recommending second user, easy to operate, and then can be dropped
The probability of the low game for recommending non-user demand or hobby to second user, improves the accuracy rate of game recommdation, while can also
The redundancy rate for reducing game recommdation, the user's viscosity for enhancing terminal.
Below in conjunction with Fig. 2 to Fig. 9 to application recommended method (such as game recommdation method) provided by the embodiments of the present application
And device is specifically described.
Embodiment one:
Application recommended method provided by the embodiments of the present application is applicable to recommend any application to any user, retouches for convenience
State, below any user can be illustrated by taking the first user as an example, any of the above-described application can by second application for carry out
Explanation.Assuming that including that a plurality of applications can be recommended to the first user in the application resource of terminal built-in, wherein in above-mentioned application resource
Including a plurality of applications be combined into the set of applications to be recommended of terminal.Wherein, per a application in set of applications to be recommended
A recommended priority can be corresponded to, and then can be recommended this plurality of application using corresponding recommended priority according to each money
The application recommendation list for recommending application to the first user can be obtained preceding in sequence, recommended priority higher application sequence.
The embodiment of the present application can belong to according to the user for the application attribute feature and the first user that money each in set of applications to be recommended is applied
Property feature determine the recommended priority of any money application in set of applications to be recommended, and then can be obtained to the first user and this recommended to answer
Used time, this applied the display position in application recommendation list.
Optionally, the application attribute feature that the embodiment of the present application can also be applied according to money each in set of applications to be recommended, with
And the user property feature of application to be recommended determines and recommends a application of certain in set of applications to be recommended (such as first application)
Each user corresponding recommended priority when to each user, so can obtain recommending first application some users (such as
Second user) when the corresponding recommended priority of second user, and then can be determined whether to second user recommend first application.It is optional
, however, it is determined that recommend the first application to second user, then can further determine that first applies the display in application recommendation list
Position.For details, reference can be made to the above-mentioned implementations for recommending application to the first user, and details are not described herein.
Referring to fig. 2, Fig. 2 is the flow diagram provided by the embodiments of the present application using recommended method.Such as Fig. 2, this Shen
Please embodiment provide game recommdation method may include following steps S201-S203:
S201 obtains the application feature of application to be recommended.
It in some possible embodiments, may include that the first attributive character and the second attribute are special in above-mentioned application feature
Sign.Wherein, above-mentioned first attributive character can be able to be the second application attribute spy for the first user property feature, the second attributive character
Sign.Alternatively, optional, above-mentioned first attributive character is the first application attribute feature, above-mentioned second attributive character is second user
Attributive character.
Optionally, game recommdation method provided by the embodiment of the present application is suitable for a certain user (such as first user)
Recommend the application such as any game (such as second application), above-mentioned first attributive character is the first user property feature, the second attribute
Feature is the second application attribute feature.Wherein, the user property of above-mentioned first user property feature concretely the first user is special
Sign can be illustrated for convenience of description by taking the user property feature (referred to as the first user property feature) of the first user as an example.On
Stating the second application attribute feature can be the application attribute feature of the second application, describe for convenience, can be with the second application attribute feature
For be illustrated.
Optionally, game recommdation method provided by the embodiment of the present application is suitable for a certain application (such as first application)
Recommend to any user (such as second user), above-mentioned first attributive character is the first application attribute feature, above-mentioned second attribute
Feature is second user attributive character.Wherein, the application attribute of above-mentioned first application attribute feature concretely the first application is special
Sign can be illustrated for convenience of description by taking the application attribute feature (referred to as the first application attribute feature) of the first application as an example.On
The user property feature that second user attributive character can be second user is stated, is described for convenience, it can be with second user attributive character
For be illustrated.
In the embodiment of the present application, in game recommdation method provided by the embodiment of the present application, by a certain application (such as the
One application) to the implementation of any user (such as second user) recommendation, recommend with to a certain user (such as first user)
The implementation of the applications such as any game (such as second application) is similar, therefore, below will be to recommend second to answer to the first user
It is illustrated for implementation.
In some possible embodiments, terminal is recommended to the first user in application, can obtain the first user's first
User attribute data (i.e. the first user attribute data), the user of the first user is determined according to the user attribute data of the first user
Attributive character (i.e. the first user property feature).Wherein, the user attribute data of above-mentioned first user can be for unique identification
The data of first user, the user attribute data of the first user include but is not limited to the first age of user, first user's gender,
Region locating for one user educational background, the first user and the first user are using account etc..Wherein, above-mentioned first user can using account
It is used to be associated with the application account of a plurality of applications to be recommended for the first user.For example, above-mentioned first user can be the using account
The wechat account of one user, above-mentioned a plurality of applications to be recommended are assumed to be a plurality of game in wechat game, then the first user can make
With some or all of game in its wechat account relating wechat game.Its wechat account can be used to swim in wechat for first user
Game role is registered in some or all of game of play, and then wechat account can be associated with game role foundation.
In some possible embodiments, terminal can be according to user group's attribute of the second application, by the second application
User group is grouped, and then the user attribute data of the first user can be combined to determine the first user couple according to the result of grouping
The the first user property feature answered.For example, terminal can be according to male user and for women in user group's attribute of the second application
The user of second application, is divided into two groupings of male user and female user by the user group's attribute having per family, and can be
A label is set separately in male user and female user, and then user attribute data can be converted to user property feature.Example
Such as, it is assumed that terminal is respectively used to label male user and female user with 1 and 0, and by the characteristic parameter of 2 characters for marking
The sex character of any user of the second application of note.For example, if the user attribute data of the first user is the gender data of male,
Terminal can then determine that the first user property feature of the first user is masculinity according to the user data attribute of the first user, into
And it can determine that the first user property feature of the first user is 10.If the user attribute data of the first user is the gender number of women
According to terminal can then determine that the first user property feature of the first user is that women is special according to the user data attribute of the first user
Sign, and then can determine that the first user property feature of the first user is 01.And so on, when the user attribute data of the first user
It is region locating for the first age of user, the first user educational background, the first user and the first user using in the data types such as account
It is any, and terminal can also be arrived according to the mode classification and the first user attribute data of above-mentioned first user attribute data
The conversion regime of first user property feature determines that corresponding first user of the first user attribute data of the various forms of expression belongs to
Property feature.It can specifically be determined according to practical application scene, herein with no restrictions.
In some possible embodiments, terminal is to the first user recommendation second in application, the second application can be obtained
Application attribute data (i.e. the second application attribute data) determine the application of the second application according to the application attribute data of the second application
Attributive character (i.e. the second application attribute feature).Wherein, above-mentioned second application attribute data may include that the application of plurality of classes belongs to
Property data, including but not limited to application identities, application type, active degree instruction information, application resource type and user's row
For data etc., herein with no restrictions.Wherein, above-mentioned application identities can be application ID, can also be to apply in a certain set of applications (example
Set of applications such as to be recommended) in number, herein with no restrictions.Above-mentioned application type includes but is not limited to leisure, chess and card, competing
Skill, role, movement, intelligence development and single machine etc., wherein the application type of the second application can be one in above-mentioned a variety of application types
Kind.Above-mentioned active degree instruction information may include online hours or in line frequency etc., wherein the active degree of the second application refers to
Show that information may include that the first user logs in the online hours of the second application, in line frequency etc..Above-mentioned application resource type may include
Free application, payment applications or application on probation etc., wherein the application resource type of the second application can be above-mentioned a variety of application moneys
One of Source Type.Above-mentioned user behavior data can for user operation habits, user charges record, user's online hours and
User's online period etc., wherein the user behavior data of above-mentioned second application can be one in above-mentioned a variety of user behavior datas
Kind is a variety of, herein with no restrictions.Optionally, the first user attribute data of the second application can be by above-mentioned plurality of classes
A kind of classification application attribute data composition can also be the composition of the plurality of classes application attribute data of above-mentioned plurality of classes, specifically
It can be according to practical application scene, herein with no restrictions.
In some possible embodiments, the second application attribute data that terminal can be applied according to second, and wait push away
The application attribute data for recommending the application to be recommended of other in set of applications are divided the second application attribute data of the second application
Group, and second is determined according to the conversion regime of group result and above-mentioned first user attribute data to the first user property feature
The corresponding second application attribute feature of the second application attribute data of application.For example, it is assumed that the second application is in wechat game
King's honor, set of applications to be recommended are the M money game in wechat game, and wherein M is the natural number greater than 2.Assuming that M money game
Type of play in include leisure, chess and card, sports and intelligence development, king's honor belongs to the sports in this 4 kinds of type of play.Terminal
The characteristic parameter that byte length is 4 characters can be used, the corresponding type of play of each money game is converted into each money game correspondence
One of game attributes feature, wherein an a kind of type of play of character representation, and respectively with " 1 " and " 0 " indicate "Yes" with
"No".For example, it is assumed that including type of play in the game attributes data of king's honor to race, terminal then can be by king's honor
For game this game attributes data of type of racing, to be converted to the corresponding game attributes feature of king's honor be 0010.
Whether the type of play of 0010 expression king's honor lies fallow (0), is not chess and card (0), is sports (1), is not intelligence development (0).It is false
If in the game attributes data of king's honor including type of play for leisure and racing, king's honor can be both then to stop by terminal
Not busy type is also that race game this game attributes data of type are converted to the corresponding game attributes feature of king's honor
It is 1010.The type of play of 1010 expression king's honors is leisure (1), is not chess and card (0), is sports (1), is not intelligence development (0),
Deng.The type of play of above-mentioned king's honor is classified and game attributes Feature Conversion mode is only citing, specifically can be according to reality
Application scenarios are determining, herein with no restrictions.Similarly, terminal can be used similar implementation and determine that other of the second application second are answered
The second application attribute feature corresponding to attribute data, details are not described herein.
Optionally, the application that above-mentioned second application can be paid close attention to or download or play before for the first user,
At this point, user behavior data in the second application attribute feature can for second apply the first user concern process or under
The user behavior data for recording and/or storing in load process or use process.At this point, the recommendation of the second application can be described as drawing
Reflux application is recommended.It retracts stream application recommendation and can be understood as one kind and be intended to call old user to pay close attention to or reuse again the
The application way of recommendation of two applications, in other words it can be appreciated that one kind is intended to draw frequent customer's (user) to apply recommendation side
Formula.It is to be understood that the first user recommend the first user before paid close attention to perhaps played however in the recent period do not notice or
The application that do not play, to cause the concern again of the first user or play.
Optionally, what above-mentioned second application can be not concerned with or download or play before this for the first user answers
With including online for a long time or just online or i.e. by the application of the types such as online.At this point, in the second application attribute feature
User behavior data then can be set to 0 for sky, corresponding user behavior characteristics.Corresponding, the recommendation of such the second application can claim
To draw new opplication to recommend, it can be understood as a kind of application way of recommendation for being intended to draw new user.It is to be understood that the first user
The application for recommending the first user always not play before this, to cause concern and/or the downloading of the first user to be played.
The application way of recommendation provided by the embodiments of the present application is both applicable to retract stream application recommendation, is equally applicable to draw new
Using recommendation, it is equally applicable to the combination of both forms or other more multi-form applications is recommended, herein with no restrictions, operation
Flexibly, applied widely.
In some possible embodiments, terminal acquires the first user property feature and the second application attribute feature
Later, then the first user property feature and the second application attribute feature can be combined to construct second and apply corresponding application
Feature.The user property feature of first user is used in the building using feature of the second application by terminal, and the second application pushes away
It recommends close with the user property characteristic relation of the first user, enhances the recommendation and the relevance of the first user of the second application, behaviour
Make feasibility simple, and then that the orientation recommendation of the second application can be enhanced, applicability is stronger.
It similarly, in the embodiment of the present application, can also be according to upper in the implementation the first application recommended to second user
The application feature for stating the first application of implementation building, including the first application attribute feature and second user attributive character,
So that the user property characteristic relation of the recommendation of the first application and second user is close, the recommendation of the first application and the is enhanced
The relevance of two users.For details, reference can be made to above-mentioned implementations, herein with no restrictions.
S202 is determined according to the first attributive character of application to be recommended and is applied recommendation network model, is recommended by the application
Network model learn to the second attributive character and determines the corresponding application recommendation of application to be recommended or application to be recommended
Corresponding user's recommendation.
In some possible embodiments, terminal acquires the application feature of application to be recommended (such as second application)
Later, then using second application using the first user property feature for including in feature from apply recommendation network model set
In include it is multiple using in recommendation network model matching obtain the first user-association application recommendation network models (for convenience of describe
Recommendation network model can be applied labeled as first).It is application recommendation side provided by the embodiments of the present application please also refer to Fig. 3, Fig. 3
The another application schematic diagram of a scenario of method.Terminal can store multiple application recommendation network model (abbreviation model, examples in its database
Such as model 1, model 2 ..., model n), wherein above-mentioned multiple models for realizing to multiple users application orientation recommend.
For convenience of description, the multiple application recommendation network models stored in above-mentioned database be can be described as using recommendation network model set,
I.e. using may include in recommendation network model set above-mentioned such as model 1, model 2 ..., the multiple application recommendation network moulds of model n
Type.Optionally, above-mentioned database may also be stored in the corresponding server end of terminal, can specifically be determined according to practical application scene,
Herein with no restrictions.Wherein, the identical network architecture can be used in each model shown in Fig. 3, and can be associated with by different user
Sample application sample application feature train obtain different network parameters.Wherein, a model can be associated with a user and belong to
Property feature, each associated user property feature of model can be used for marking the model to be applicable to realize to user property spy
That levies corresponding user applies recommendation function.As shown in figure 3, model 1, model 2 ..., model n etc. identical network rack can be used
Structure, wherein the sample application feature training that model 1 can be applied by the associated sample of user 1 obtains a set of network parameters, so that tool
The model 1 of the standby set of network parameters can be used for realizing that the application orientation to user 1 is recommended.Likewise, model 2 can be closed by user 2
The sample application feature training of the sample application of connection obtains a set of network parameters, so that the model 2 for having the set of network parameters can
Recommend for realizing the application orientation to user 1.Optionally, above-mentioned model 1, model 2 ..., model n etc. can also answer to be same
With recommendation network model, and there can be sample application the characteristic of the associated sample application of different user to train and obtain different nets
Network parameter realizes that the application orientation to different user is recommended with this, can specifically be determined according to practical application scene, not limited herein
System.
Optionally, terminal can store in its database multiple application recommendation network models (abbreviation model, for example, model 1,
Model 2 ..., model n etc.), wherein above-mentioned multiple models are recommended for realizing orienting to the user of multiple applications.Wherein, Fig. 3
Shown in each model can be used the identical network architecture, and can be special by the sample of users of the associated sample of users of different application
Sign training obtains different network parameters.Wherein, a model can be associated with an application attribute feature, and each model is associated
The application recommendation that application attribute feature can be used for that the model is marked to be applicable to realize to the corresponding application of application attribute feature
Function.As shown in figure 3, model 1, model 2 ..., model n etc. the identical network architecture can be used, wherein model 1 can be by user 1
The sample of users feature training of associated sample of users obtains a set of network parameters, so that having the model 1 of the set of network parameters
Can be used for realizing the application orientation of application 1 corresponding to sample of users recommend (one or more user will be recommended using 1,
Or the directional user of application 1 recommends).Likewise, model 2 can be instructed by the sample of users feature of 2 associated sample of users of application
It gets to a set of network parameters, so that the model 2 for having the set of network parameters can be used for realizing that the application orientation to application 1 is recommended
(one or more user will be recommended using 2, or the directional user of application 2 recommends).Optionally, above-mentioned model 1, model
2 ..., model n etc. can also be the same application recommendation network model, and can have the sample of the associated sample of users of different application to use
Characteristic training in family obtains different network parameters, realizes directional user's recommendation to different application with this, specifically can basis
Practical application scene is determining, herein with no restrictions.Below by the model to recommend for realizing the application orientation to different user
For be illustrated.
In some possible embodiments, when realizing that the application orientation to user 1 (can be set as the first user) is recommended,
It can using the user property feature for adding user 1 in feature, (such as user belongs in any application to be recommended for recommending to user 1
Property feature 1), and then can be matched from multiple models by user property feature 1 and obtain model 1.And then it can be by application to be recommended
Application feature input model 1, and then the recommendation that application to be recommended be recommended to user 1 can be exported by model 1.If wait push away
It recommends using being multiple applications, then passes through the then corresponding recommendation of exportable each application of model 1, and then can be according to the big of recommendation
The recommended priority of the small each application of determination generates and applies recommendation list.It will be carried out by taking the first application to be recommended as an example below
Explanation.
In some possible embodiments, it may also include the application respectively applied to be recommended in database shown in Fig. 3
The user attribute data of attribute data and multiple users.As shown in figure 3, when the user clicks the application functions such as game when, terminal
The user operation instruction on the corresponding operating areas of application functions such as game can be detected, and then produce and apply recommendation request,
Wherein, this is using the user attribute data that can carry the first user in recommendation request.After recommendation request is applied in terminal generation, then
The multiple applications stored in database can be called according to the corresponding first user property feature of user attribute data of the first user
First in recommendation network model applies recommendation network model.For example, if the first user is user 1, it can be by the user of user 1
Attributive character and application recommendation network model set in include each user property feature using recommendation network model interaction into
The corresponding application recommendation network mould of user property feature of user 1 is determined in row matching from application recommendation network model set
Type, such as model 1.The user property feature as associated by model 1 is the user property feature of user 1, and model 1 can also
It is equivalent to the associated application recommendation network model of user 1, herein with no restrictions.
It in some possible embodiments, then can be according to after terminal acquires the application feature of the second application
The application recommendation network model that the first user-association is called using the first user property feature for including in feature of two applications, and
The application feature of second application is inputted this using recommendation network model, is applied using recommendation network model to second by this
The recommendation for being learnt using the second application attribute feature in feature, and being applied based on application recommendation network model prediction second
Second application is recommended the first user by value.Wherein, the second application can be one in multiple applications to be recommended, therefore the
Two applications can be exported using recommendation list to the user interface of terminal with other application compositions to be recommended, such as the interface in Fig. 3
3.In the specific implementation, any one in multiple application recommendation network models as shown in Figure 3 can be by using recommendation network model
Sample application the feature of a large amount of sample application, which is trained, to be obtained, and it is each using recommendation network model by user-association
Sample application training obtains, and then each application orientation for being adapted to carry out a user using recommendation network model may make to push away
It recommends.For convenience of description, will be illustrated by taking the corresponding application recommendation network model of the first user as an example below.Above-mentioned first user
It is corresponding to be obtained by the training of a large amount of sample application feature of the first user-association using recommendation network model and above-mentioned big
It include sample application attribute feature and the first use of above-mentioned first user in each sample application feature in the sample application of amount
Family attributive character.Wherein, above-mentioned sample application can be used a plurality of applications before the first user at the appointed time t, and every
It can be associated with by the foundation of the first user attribute data of the first user with the first user when money is applied as sample application, into
And it can be used for the recommendation for any application that training is recommended using recommendation network model for pre- the first user of direction finding.
In some possible embodiments, the sample application feature of the corresponding a plurality of sample applications of the first user can be used as
Using the input of recommendation network model, predict that the application feature of each sample application is corresponding by application recommendation network model
Sample application recommendation is that learning tasks learn the application feature that each sample is applied.The sample applied by a plurality of samples
The study of this application feature can carry out repetition training to the recommendation forecast function of application recommendation network model, and then may make and answer
There is any using the corresponding ability using recommendation of feature of prediction input with recommendation network model.Terminal is using trained
To when predicted using recommendation using recommendation network model the second of the second application, can be special by the application of the second application
Sign as apply recommendation network model input, and then can by using recommendation network model to second apply application feature into
Row learns and exports the application feature corresponding second of the second application using recommendation.
Optionally, in some possible embodiments, the sample of users of the first corresponding multiple sample of users of application is special
Sign can be used as the input using recommendation network model, by application recommendation network model to predict that the user of each sample of users is special
Levying corresponding sample of users recommendation is that learning tasks learn the user characteristics of each sample of users.Pass through multiple samples
The study of the sample of users feature of user can carry out repetition training to the recommendation forecast function of application recommendation network model, in turn
It may make the ability for the corresponding user's recommendation of any user feature that there is prediction input using recommendation network model.Terminal benefit
With training obtain when being predicted using second user recommendation of the recommendation network model to second user, can be by second user
User characteristics as the input for applying recommendation network model, and then can pass through using recommendation network model to the use of second user
Family feature is learnt and exports the corresponding second user recommendation of application feature of second user.Wherein, above-mentioned second user
The recommendation of second user when recommendation can be to recommend second user for the first application, and then can recommend according to by the first application
To the comparison of the recommendation of the other users except second user, it is determined whether recommend the first application to second user.Specifically may be used
Referring to above embodiment, details are not described herein.
S203 recommends the application of application to be recommended preferential according to determining using recommendation to the first user for application to be recommended
Grade, or the User Priority for recommending the first application is determined according to user's recommendation of application to be recommended.
In some possible embodiments, terminal has been determined to the first user recommendation second in application, the second application pushes away
After the second application recommendation recommended, then it can determine that second applies in set of applications to be recommended using recommendation according to second
Recommended priority.Wherein, the second application recommendation is bigger, and second applies the recommended priority in set of applications to be recommended then to get over
It is high.It is corresponding, second apply in set to be recommended all applications to be recommended compositions using the recommendation order in recommendation list
It is then more forward.When the application priority of the second application, which is more than or equal to, to be preset using priority threshold value, pushed away to the first user
Recommend the second application.
For example, by taking this kind of application of game as an example, please also refer to Fig. 1, it is assumed that recommend to the first user in application, to be recommended
Include in set of applications game 1, game 2 ..., a plurality of game such as game 5, by above-mentioned steps S201 into S203 each step
Described implementation, which can be realized, recommends list (" I " as shown in figure 1) for the personal of first user's recommended games.Assuming that
Second application described in above-mentioned each step is game 3, and the corresponding recommendation of game 3 is higher than in set of applications to be recommended
Game 3 can then be sorted in the personal umber one for recommending list of the first user, and used for first by the recommendation of other each money game
Family using first in recommendation list.First user is in addition to consulting the trip of heat listed by the user group for each money game
After list, good friend's heat play the popular recommendation information such as list and new trip list, each money game released for me can be also checked
Exclusive ranking list, and then the game oneself liked can be quickly found out.
Optionally, in the embodiment of the present application, after terminal has determined the second user recommendation that second user is recommended, then
Recommended priority of the second user in user to be recommended set can be determined according to second user recommendation.Wherein, second user
Recommendation is bigger, and recommended priority of the second user in user to be recommended gathers is then higher, and corresponding, second user is wait push away
The recommendation order recommended in set in user's recommendation list of all user's compositions to be recommended is then more forward.As the user of second user
When priority is more than or equal to pre-set user priority threshold value, determines to second user and recommend the first application, to can realize
Second user is recommended using orientation by first, the user that the application of the first application of enhancing is recommended is efficient.
In the embodiment of the present application, the user property feature of the first user can be used for the application feature of the second application by terminal
Building in, and then can by the user property feature of the first user and second apply application attribute feature constructed by obtain answer
It is sent into feature using in recommendation network model, is learnt by application feature of the application recommendation network model to the second application
With the recommendation of the second application when pre- the first user of direction finding the second application of recommendation, and then it can realize that the orientation to the first user pushes away
It recommends, it is easy to operate.Further, in the embodiment of the present application, the user property feature of second user can be used for first by terminal
In the building using feature of application, and then the application attribute feature that the user property feature of second user and first can be applied
Constructed obtained application feature is sent into using in recommendation network model, and the use using recommendation network model to second user is passed through
Family feature is learnt to recommend with pre- direction finding second user user's recommendation when the first application, and then can be realized the first application
Orientation recommends second user.The user property feature of first user is dissolved into the recommendation of the second application by the embodiment of the present application
Cheng Zhong, and/or the application attribute feature of the first application is dissolved into the application recommendation process of the first user, it enhances application and pushes away
The affinity that is associated with user is recommended, and then the probability for recommending the game of non-user demand or hobby to user can be reduced, is improved
Using the accuracy rate of recommendation, while the user's viscosity for enhancing terminal using the redundancy rate recommended can also be reduced.
Embodiment two:
In application recommended method provided by the embodiments of the present application, terminal is by applying recommendation network model realization to first
User recommends application, and/or specified by recommending second user etc. using orientation for first using recommendation network model realization
The implementation complexity that can be also substantially reduced using the accuracy rate recommended using recommending not only can be improved in user group, improves application
The efficiency of recommendation.It greatly reduces using the addition of recommendation network model using the data processing complexity recommended, while also big
The accuracy and feasibility recommended using orientation are improved greatly.Below in conjunction with Fig. 4 and Fig. 5, to provided by the embodiments of the present application
It is illustrated using the building process of recommendation network model.For convenience of description, recommend net in application provided by the embodiments of the present application
It will be the first user property feature, the second attributive character with the first attributive character be the second application in the building process of network model
It is illustrated for attributive character.It is appreciated that the realization of the building provided by the embodiments of the present application using recommendation network model
Mode is applied equally to that the first attributive character is the first application attribute feature, the second attributive character is second user attributive character
Recommendation application recommendation, herein with no restrictions.
It referring to fig. 4, is another flow diagram provided by the embodiments of the present application using recommended method.Implement in the application
In example, the building using recommendation network model may include three phases, including stage one: sample characteristics building;Stage two: network
Model training and stage three: network model test.
Stage one: sample characteristics building may include following steps S401-S403.
S401 obtains the sample data for recommending at least two samples application of training for application.
S402 constructs the corresponding sample characteristics of sample data of various kinds this application in above-mentioned at least two samples application.
In the specific implementation, including the in the sample data of any one sample application in the application of above-mentioned at least two sample
The sample application attribute data of the first user attribute data and the sample application of one user, corresponding, any of the above-described sample application
Sample characteristics in include the first user property feature and sample application attribute feature.
In some possible embodiments, terminal, which is obtained, recommends the sample of training for application in application, the can be chosen
One user before nearest time t a plurality of applications.For example, when choosing the sample game of game recommdation training, terminal
The first user can be chosen before nearest time t in a plurality of game of object for appreciation.Terminal has chosen a plurality of applications as sample using it
Afterwards, then the application attribute data of the application per a sample in this plurality of sample application be can record, i.e., application corresponds to per a sample
Sample application attribute data.It is appreciated that in order to guarantee using recommendation network model to any application using recommendation
Forecasting accuracy, in the service stage and training stage of application recommendation network model, input is each using recommendation network model
The building mode using feature of money application should be consistent.Therefore, corresponding sample application is applied for every a sample
The record of attribute data can record per a sample the application identities of application, application type, active degree instruction information, application
One or more of a variety of application attribute data categories such as resource type and user behavior data.Wherein, above-mentioned a variety of
The form of expression of the application attribute data of various application attribute data categories can be found in above-mentioned steps in application attribute data category
It is described accordingly in S201, details are not described herein.
In a kind of feasible embodiment, it is assumed that application attribute data (the i.e. sample application attribute of each sample application
Data) in include the multinomial attribute data such as application identities, application type and user behavior data.Assuming that the sum of sample application
For M, then above-mentioned M sample application can be numbered to obtain numbering the M sample application for 1-M.It is corresponding, it is above-mentioned each
The application identities of sample application can apply the number in M sample application then for each sample.It is please also refer to Fig. 5, Fig. 5
Building schematic diagram provided by the embodiments of the present application using recommendation network model.In the stage 1 shown in Fig. 5, it is assumed that M sample
It include numbering the sample for being 2 to apply 5 using the sample that 2 and number are 5 in.Wherein, sample is using 2 and sample using 5
It include application identities (i.e. sample applies 2 number 2 and the number 5 of sample application) in sample application attribute data, using class
Type and user behavior data.The application identities that each sample is applied can be normalized for terminal, i.e., answer each sample
Application identities are converted into the application identities feature of same expression way.For example, the application identities that each sample is applied
It is transformed to nondimensional application attribute feature from the expression formula for having dimension, each sample is answered using recommendation network model with reducing
The study complexity of application attribute feature.As shown in figure 5, terminal can obtain the application identities that each sample is applied divided by M
The corresponding application identities feature of application identities identified to each money.Wherein, normalized mode shown in fig. 5 is only that one kind is shown
Example can specifically determine, herein with no restrictions according to practical application scene.
As shown in figure 5, above-mentioned application type can the conversion regime as shown in above-described embodiment one, by sample apply 2 sample
The application type attribute data for including in this application attribute data is converted to sample using 2 corresponding application type features
1001.Similarly, sample sample can be converted to using the application type attribute data for including in 5 sample application attribute data to answer
It is 0110 with 5 corresponding application type features.For details, reference can be made to the application attribute data of the offer of embodiment one to using characteristic
According to conversion regime, details are not described herein.
In some possible embodiments, as shown in figure 5, what sample was applied using 2 and sample using each samples such as 5
User behavior data is further comprised in application attribute data.Corresponding user behavior data is applied in above-described embodiment one second
Type it is identical, the user behavior data of each sample application may also comprise login state, user operation habits, user charges note
The data of many forms such as record, user's online hours and user's online period.For example, sample shown in fig. 5 applies 2 Hes
Sample is login state, user operation habits, user charges note in 5 application attribute data using the user behavior data for including
4 user behavior parameters such as record and user's online hours.Wherein login state includes registered users and non-registered users etc., and
" 1 " and " 0 " can be used respectively to be marked.Above-mentioned user operation habits include user's self-defining operation mode and system definition behaviour
Make mode, and " 1 " and " 0 " is respectively adopted and is marked.Above-mentioned user charges record includes that user's charges paid and user are non-paid,
And " 1 " and " 0 " is respectively adopted and is marked.Above-mentioned user's online hours include online hours be less than or equal to X hour with
Online hours are greater than X hour, and " 1 " and " 0 " is respectively adopted and is marked, and X is the natural number greater than 0.Terminal determines
Characteristic parameter to mark user behavior data is the characteristic parameter of 4 characters, and determines what each character was marked
After user behavior parameter, then can according to each sample apply application attribute data in include user behavior data, will be each
User behavior data in the application attribute data of a sample application is converted to the application attribute feature of sample application.For example, sample
User behavior data in the sample application attribute data of this application 2 is registered users, system defining operation mode, Yong Huyi
Payment and user's online hours are greater than X hour, then sample can be converted to sample using 2 user behavior data and answered by terminal
It is 1011 with the user behavior characteristics in 2 sample application attribute feature.Similarly, sample can be applied 5 user behavior by terminal
Data are converted to sample using the user behavior characteristics 0111 in 5 sample application attribute feature, and details are not described herein.
In the specific implementation, recommending to realize to the application of the orientation of the first user, sample of the terminal in building sample application
When application attribute feature, it is also necessary to which the first user property feature of the first user is added to the sample application of each sample application
In attributive character.As shown in fig. 5, it is assumed that sample is using 2 and sample using in the user attribute data for including in 5 sample data
It include this attribute data of the first age of user, terminal then can be according to the age distribution etc. of the user group of various kinds this application
The age of user group is segmented so that each user to be grouped by information.For example, if various kinds this application user of interest
Group is user of the age distribution between 18 years old to 40 years old, then user group can be divided into the user group of 4 age brackets, is wrapped
Include 18 years old to 25 years old one group (group 1 can be set for convenience of description), 25 years old to 30 years old one group (group 2), the one of 30 years old to 35 years old
Group (group 3) and one group of 35 years old to 40 years old (group 4).Terminal can by byte length be 4 characters age characteristics parameter come
Indicate the age characteristics of the first user, wherein 4 characters mark above-mentioned 4 ages to be grouped respectively.For example, if first user
Age is 20 years old, then can determine that the age of the first user is included in group 1, and can be existed with the age of the first user of one token
In group 1.Wherein, group 2, group 3 and group 4 can be used " 0 " and be marked, indicate the age of the first user not in the grouping, into
And the corresponding first user property feature of user attribute data (i.e. age) of available first user is " 1000 ".With such
It pushes away, if the user attribute data of the first user is the data of other forms of expression except the age, similar realization can also be used
The user attribute data of first user is converted to the corresponding user property feature of the first user by mode, specifically can be according to actually answering
Determine that details are not described herein with scene.As shown in figure 5, applying 5 using 2 and sample as sample is associated by the first user
Sample application, therefore, sample is using 2 and sample using including the first user corresponding the in 5 corresponding sample characteristics
One user property feature, i.e. " 1000 ".
In the embodiment of the present application, terminal has determined the first user property feature and various kinds this application of the first user
Sample application attribute feature after, then can by the sample application attribute feature of the first user property feature and various kinds this application into
Row combination, obtains the corresponding sample characteristics of various kinds this application.Sample as shown in Figure 5 applies 2 corresponding sample characteristics " 1000
2/M 1,001 1011 " and sample apply 5 corresponding sample characteristics " 1000 5/M 0,110 0111 ".
S403 constructs at least one sample application feature pair according to the sample data that above-mentioned at least two sample is applied.
In some possible embodiments, in order to preferably from multiple samples application in identify the first user more loading
The application of interest, the mode matched two-by-two between multiple sample applications can be used in the embodiment of the present application, from two sample applications
In elect the more interested application of the first user, and then the application recommended models obtained by sample application training can be improved
Recommendation accuracy rate.In the specific implementation, terminal can be more to obtain using be matched two-by-two by M sample of the first user-association
A sample application pair.Further, it is applied according to the application attribute data of each sample application from each sample and is determined in
Positive sample is applied and negative sample application.For example, the active degree that terminal can be applied according to each sample indicates information from each sample
Positive sample application and negative sample application are determined in two samples application of this application pair, wherein positive sample application enlivens journey
Degree is higher than the active degree of negative sample application.Alternatively, terminal can be according to user group's quantity that each sample is applied from each sample
Positive sample application and negative sample application are determined in two samples application of this application pair, wherein the user group of positive sample application
Body quantity is greater than user group's quantity etc. of negative sample application.Wherein, the sample application attribute that terminal is applied according to each sample
The implementation that data determine that positive sample application and negative sample are applied from two samples application of each sample application pair can
It is determined according to practical application scene, herein with no restrictions.For convenience of description, active degree instruction information will be detailed below as positive sample
It is illustrated for this application and the division parameter of negative sample application.
In some possible embodiments, terminal can be applied according to any one sample to (for example sample is applied to i)
In two samples application active degree instruction information determine positive sample apply and negative sample application.Such as Fig. 5, it is assumed that sample
Applied in 5 sample application attribute data using 2 and sample includes that active degree indicates information, and applies 2 by sample
It can determine that sample is enlivened using 2 using the active degree instruction information for including in 5 sample application attribute data with sample
Degree is higher than sample and applies 5, then sample can be determined as positive sample application using 2, sample is determined as negative sample using 5 and is answered
With.Corresponding, the sample characteristics of positive sample application are then positive sample characteristics (for convenience of describing that positive sample feature i can be labeled as1),
Then negative sample feature (can be labeled as positive sample feature i for convenience of description to the sample characteristics of negative sample application0), and then sample can be obtained
This application is to the corresponding sample application feature of i to i10.That is, a sample application feature centering includes a positive sample feature and one
A negative sample feature, also, corresponding in positive sample feature includes the first user property feature and positive sample application attribute feature, is born
Corresponding in sample characteristics includes the first user property feature and negative sample application attribute feature.
In some possible embodiments, terminal has determined that various kinds this application centering includes positive sample feature and negative sample
After feature, then it can be used to input the sample using training in recommendation network model according to positive sample feature and negative sample latent structure
This application feature pair, tectonic style can are as follows: (pair<sample_p, sample_n>, 1).Wherein, sample_p indicates positive sample
Feature, sample_n indicate negative sample feature, and 1 is the label (label) for marking positive sample feature.(pair<sample_
P, sample_n >, 1) it indicates for the first user, to the application of sample corresponding to positive sample feature, (sample is applied
ID can be set as interest level p) and be greater than the application of sample corresponding to negative sample feature (ID of sample application can be set as n), because
The corresponding sample characteristics of this sample application p come before the corresponding sample characteristics of sample application n.
Stage two: network model training may include following steps S404-S405.
S404, by the positive sample feature of at least one above-mentioned sample application feature centering various kinds this application feature pair and negative sample
Recommendation network model is applied in eigen input.
S405 passes through the above-mentioned positive sample feature using recommendation network model to above-mentioned various kinds this application feature pair and negative sample
Eigen is learnt.
In some possible embodiments, terminal constructs at least one sample application feature to later, then can will be each
The positive sample feature and negative sample feature of sample application feature centering are input to using in recommendation network model, are recommended by application
Network model learns the positive sample feature and negative sample feature of various kinds this application feature pair, and then can make using recommendation net
It is any using the corresponding ability using recommendation of feature that network model obtains prediction.Wherein, any sample application feature centering is being just
The corresponding application recommendation of sample characteristics is corresponding greater than negative sample feature to apply recommendation, i.e. the recommendation of positive sample application is preferential
Grade is higher than the recommended priority of negative sample application.Please also refer to Fig. 5, in the embodiment of the present application, using recommendation network model
Sorting network model can be used.It since what is inputted when application recommendation network model training is sample application feature pair, that is, include two
A sample characteristics, and the input of two sample characteristics is the same network model, therefore for convenience of describing, mirror image mould can be used
The mode of type is illustrated the training process provided by the embodiments of the present application using recommendation network model.
In some possible embodiments, such as Fig. 5, it is assumed that include two modules, two moulds using recommendation network model
Model structure in block is completely the same, and the network parameter (such as activation primitive) of model is also completely the same.By sample characteristics centering
Positive sample feature and negative sample feature when being separately input to using recommendation network model, it is equivalent with by positive sample feature and negative sample
Eigen inputs two modules respectively, obtains the corresponding recommendation of positive sample feature (being assumed to be first using recommendation) with processing
Recommendation (being assumed to be second using recommendation) corresponding with negative sample feature.Specifically, above-mentioned application recommendation network model (figure
Any module shown in 5) network structure design it is as follows:
The above-mentioned network structure using recommendation network model may be designed as include three full articulamentums network structure, specifically
It can also be the network structure of more full articulamentums, or the network structure of less full articulamentum in realization, herein with no restrictions.Its
In, if full articulamentum quantity excessively if may cause that data processing complexity is excessive so that the prediction knot of network model
Fruit handles the data processing problems such as over-fitting.If the quantity of full articulamentum is very few, the precision that may cause data processing is inadequate,
So that there are the data processing problems such as poor fitting in the prediction result of network model.Therefore, the quantity of full articulamentum can basis
The application demands such as precision of prediction of the size of data processing amount or network model determine in practical application request, do not limit herein
System.
For convenience of description, will be illustrated by taking four full articulamentums as an example below, such as Fig. 5:
First layer is full articulamentum, including N number of neuron (assuming that neuron number be 256), and each nerve
The activation primitive of member is line rectification function (rectified linear unit, Relu).The full articulamentum of first layer can lead to
Included each neuron is crossed to learn, Jin Erke the sample characteristics (positive sample feature or negative sample feature) of input
The sample characteristics obtained after study are exported to next layer of neuron and are handled.
The second layer is also full articulamentum, including N/2 neuron (i.e. neuron number be 128), and each mind
Activation primitive through member is also relu.In the network structure of model, next full articulamentum of layer is in the sample characteristics to input
Dimensionality reduction is carried out to reduce redundancy feature to the sample characteristics of upper one layer of output while study.It is appreciated that next time
What full articulamentum can be understood as exporting upper one layer full articulamentum to the output feature progress dimensionality reduction of upper one layer full articulamentum
High dimensional feature is low-dimensional feature by the mode conversions such as Feature Mapping, i.e., goes the feature of redundancies some in high dimensional feature divided by obtaining
The less feature of data volume is obtained, and then data processing complexity can be reduced while retaining essential feature.Similarly, second
Full articulamentum can be by included each neuron to sample characteristics (the positive sample feature of the full articulamentum of first layer output
Or negative sample feature) learnt, and then can will after study obtained sample characteristics export to next layer of neuron into
Row processing.
Third layer is also full articulamentum, including N/4 neuron (i.e. neuron number be 64), and each nerve
The activation primitive of member is also relu.The last layer is the full articulamentum of 1 neuron, activation primitive sigmoid.Wherein,
The full articulamentum of last time uses sigmoid function, can export by the export-restriction of model between [0,1] as model
Score (i.e. recommendation).
In some possible embodiments, special using positive sample of the recommendation network model to the sample characteristics centering of input
The corresponding recommendation of exportable positive sample feature after sign is learnt, such as first using recommendation.Similarly, using recommendation net
Network model corresponding recommendation of exportable negative sample feature after being learnt to negative sample feature, such as the second application are recommended
Value.Can also data processing be carried out to the first application recommendation and the second application recommendation using recommendation network model, and combinable
The label that sample characteristics centering carries is modified output result and then can continually strengthen the prediction using recommendation network model
The precision of recommendation improves the accuracy rate of the output recommendation using recommendation network model.In the embodiment of the present application, positive sample
The corresponding recommendation of feature may be set to 1, and the corresponding recommendation of negative sample feature is also set to 0, therefore, above-mentioned sample characteristics pair
The label carried when input can be used to the amount of the default recommendation of label positive sample feature, it can also be used to mark positive sample feature
The difference of recommendation corresponding with negative sample feature.It can specifically be determined according to practical application scene, herein with no restrictions.
In some possible embodiments, terminal can be set any in the training process of application recommendation network model
The corresponding first default recommendation (such as 1) of sample application centering positive sample application, the corresponding second default recommendation of negative sample application
(such as 0).By the positive sample feature of any sample application feature pair and negative sample feature input using recommendation network model it
Afterwards, the positive sample feature corresponding first using the output of recommendation network model can be obtained using recommendation and negative sample feature pair
Second answered applies recommendation.Terminal can be pre- in conjunction with first according to the difference of the first application recommendation and the second application recommendation
If the difference of recommendation and the second default recommendation calculates recommendation loss.For example, terminal can be by the first application recommendation and the
Two carry out making the difference the difference calculated to obtain two using recommendation using recommendation, and then difference can be carried out sigmoid change
It changes and zooms between [0,1], and the label of the output of sigmoid transformation and sample characteristics centering is subjected to costing bio disturbance to obtain
Recommendation loss.The loss of above-mentioned recommendation can be fed back using recommendation network model in the learning process of other sample characteristics,
Above-mentioned recommendation value parameter can be used for optimizing applies recommendation network model to it using the activation primitive of recommendation network model to correct
The corresponding precision of prediction using recommendation of his sample characteristics so that using recommendation network model for sample characteristics centering just
The output recommendation of sample characteristics is close in 1.I.e. so that corresponding each sample characteristics are for, using recommendation network
Model is as big as possible in negative sample spy to the application recommendation (first applies recommendation) exported after positive sample feature learning
The application recommendation that sign exports after being learnt (second applies recommendation).Repeatedly, it can constantly correct using recommendation net
The predictablity rate of network model, and then may make and have the corresponding recommendation of any application feature of prediction using recommendation network model
Ability.
Stage three: network model test may include following steps S406.
S406 is obtained and is recommended the test feature of test for application and input using recommendation network model.
In some possible embodiments, terminal training is applied after recommendation network model, can also obtain application
Recommend the application attribute data of the test application of test, and then test spy can be obtained using data processing according to test application
Sign.Wherein, the above-mentioned test application for application recommendation test can use after time t a or more for the first user
Money application, herein with no restrictions.In the specific implementation, after terminal acquires any test application, then it can be according to test application
Application attribute data processing obtain the corresponding application attribute feature of test application.Wherein, from the application attribute number of test application
It can be found in sample using corresponding implementation according to the conversion regime of the application attribute feature to test application, no longer go to live in the household of one's in-laws on getting married herein
It states.
Please also refer to Fig. 5, terminal obtains that test application is corresponding to answer according to the application attribute data processing of test application
After attributive character, then it is special the corresponding application of test application to be obtained in combination with the first user property feature construction of the first user
Sign, 3 corresponding application feature " 1000 3/M 10110101 " of test application as shown in Figure 5.Terminal acquires test application
After 3 corresponding application features, then this can be input to using feature using recommendation network model, by applying recommendation network
Model exports this using the corresponding recommendation of feature.For convenience of description, reference can be made to Fig. 5, the corresponding application feature of test application 3 can
Any module (essence is the same module) in two modules shown in fig. 5 using recommendation network model is inputted, by answering
After exporting corresponding recommendation with recommendation network model, terminal can be according to user's interest level of test application 3 to application
The recommendation of recommendation network model output is judged.If it is determined that using recommendation network model output recommendation meet reality
Border application scenarios demand then can determine that can satisfy practical application scene using the precision of prediction of recommendation network model needs
It asks, this is recommended to use scene using the application that recommendation network model then can be used for any application.If it is determined that application recommend net
The recommendation of network model output does not meet practical application scene demand, then needs to reacquire more samples and push away using to application
It recommends network model and carries out more training until acquire higher predictablity rate applies recommendation network model.
In the embodiment of the present application, terminal can by a large amount of sample apply to using recommendation network model be trained with
Building one has pre- the first user of direction finding and orients the ability for recommending the recommendation of any application.By applying recommendation network model
It realizes to the first user and recommends to apply, the realization that can be also substantially reduced using the accuracy rate recommended using recommending not only can be improved
Complexity, improves the efficiency that application is recommended, and applicability is stronger.
Embodiment three:
It is another flow diagram provided by the embodiments of the present application using recommended method referring to Fig. 6, Fig. 6.In the application
In embodiment, terminal training is applied after recommendation network model, then can will be used for using recommendation network model any wait push away
It recommends the application feature learning of application (such as second application) and exports the corresponding recommendation of application to be recommended.Wherein, to it is any to
The recommendation process for recommending application may include following steps S601-S604:
S601 obtains the application feature of application to be recommended.
The application feature input of application to be recommended is applied recommendation network model by S602.
S603, by determining that above-mentioned application feature is corresponding using recommendation using recommendation network model.
S604 above-mentioned to be recommended applies the recommendation in set of applications to be recommended preferential according to above-mentioned determine using recommendation
Grade, and export the application recommendation list of the first user.
In the specific implementation, above-mentioned steps S601 implementation performed by each step into S604 can be found in above-mentioned implementation
Implementation described in each step in example one, details are not described herein.
That is, in some possible embodiments, utilizing the trained application recommendation network model in front when using recommendation
The interested application of first user predicted, the output score using recommendation network model (recommend by i.e. corresponding application
Value) the first user is represented to the interest level of the application, and then the first user can be felt according to above-mentioned application recommendation
The application of interest is recommended.
When in some possible embodiments, using recommending, for draw it is new recommend, can be acquired first for the
The new opplication recommendation list that one user recommends.Due to respectively applying no user behavior data in new opplication recommendation list, then in structure
User behavior characteristics in the corresponding application attribute feature of each application can be set when building the application feature that application to be recommended is recommended
Zero processing.Application according to the application attribute feature of application to be recommended and the first user property feature construction application to be recommended is special
It levies and is input to using in recommendation network model, obtaining user to the recommendation of each application to be recommended.Further, terminal can
To the corresponding recommendation of each application to be recommended, the recommendation of each application to be recommended is determined in the way of from high to low by recommendation
Priority, wherein recommendation is higher, and corresponding recommended priority is higher.Terminal can be pushed away according to each application to be recommended is corresponding
It recommends value to be ranked up each application to be recommended to obtain the application recommendation list for the first user, as shown in Figure 1 first
User-specific ranking list (i.e. " I ").
In some possible embodiments, recommend for retracting stream, deposited before user in each application to be recommended
Concern or used user behavior data, therefore, the application feature construction of corresponding each application to be recommended can be according to above-mentioned
Feature construction mode described in embodiment one and/or embodiment two executes, and details are not described herein.Likewise, building obtains respectively
After the application feature of a application to be recommended, then the application feature of each application to be recommended can be inputted and apply recommendation network mould
Type by determining that each application to be recommended is corresponding using recommendation using recommendation network model, and then can determine each wait push away
The recommended priority of application is recommended, and exports the application recommendation list of the first user.
In the embodiment of the present application, the user property feature of the first user can be used for the application feature of the second application by terminal
Building in, and then can by the user property feature of the first user and second apply application attribute feature constructed by obtain answer
It is sent into feature using in recommendation network model, is learnt by application feature of the application recommendation network model to the second application
Recommend recommendation when the second application with pre- the first user of direction finding, and then can realize the orientation recommendation to the first user, operation letter
It is single.The user property feature of first user is dissolved into the recommendation process of the second application by the embodiment of the present application, enhances application
Recommend with the affinity that is associated with of user, and then can reduce to user recommendation non-user demand or hobby game probability, mention
The accuracy rate that height application is recommended, while the user's viscosity for enhancing terminal using the redundancy rate recommended can also be reduced.
Example IV:
It is appreciated that the implementation of the building provided by the embodiments of the present application using recommendation network model is equally applicable
In pushing away for the recommendation application that the first attributive character is the first application attribute feature, the second attributive character is second user attributive character
It recommends, herein with no restrictions.
It is another flow diagram provided by the embodiments of the present application using recommended method referring to Fig. 7.Implement in the application
In example, the building using recommendation network model equally may include three phases, including stage one: sample characteristics building;Stage two:
Network model training and stage three: network model test.
Stage one: sample characteristics building may include following steps S701-S703.
S701 obtains the user data for recommending at least two sample of users of training for application.
It in some possible embodiments, include above-mentioned first application in the user data of any of the above-described sample of users
Attribute data and sample of users attribute data.
S702 constructs the corresponding sample characteristics of sample data of each sample of users in above-mentioned at least two sample of users.
S703 constructs at least one sample of users feature pair according to the user data of above-mentioned at least two sample of users.
In some possible embodiments, a sample of users feature centering includes that a positive sample feature and one are negative
Sample characteristics, wherein in above-mentioned positive sample feature include above-mentioned first application attribute feature and positive sample user property feature, on
State includes above-mentioned first application attribute feature and negative sample user property feature in negative sample feature.
It in some possible embodiments, include active degree in the sample of users attribute data of above-mentioned each sample of users
Indicate information;
The above-mentioned user data according to above-mentioned at least two sample of users constructs at least one sample of users feature to including:
Above-mentioned at least two sample of users is matched two-by-two, obtains at least one sample of users pair;
At least one above-mentioned any sample of users of sample of users centering performs the following operations i to obtain at least one sample
User characteristics pair:
According to above-mentioned sample of users to the active degrees of two sample of users of i instruction information determine positive sample user and
Negative sample user, wherein the active degree of above-mentioned positive sample user is higher than the active degree of above-mentioned negative sample user;
Positive sample is constructed according to the sample of users attribute data of above-mentioned first application attribute data and above-mentioned positive sample user
Feature i1, and negative sample is constructed according to the sample of users attribute data of above-mentioned first application attribute data and above-mentioned negative sample user
Feature i0, above-mentioned sample of users is obtained to the corresponding sample of users feature of i to i10;
Wherein, above-mentioned sample of users feature is to i10In include above-mentioned positive sample feature i1With above-mentioned negative sample feature i0。
In the specific implementation, the sample of the above-mentioned corresponding user of the first application of user data building by least two sample of users
The implementation of feature, reference can be made to by extremely in the implementation that step S401 to S403 is provided in above-mentioned embodiment shown in Fig. 4
The implementation of the sample characteristics that the corresponding application of the first user is constructed using data of few two samples application, it is no longer superfluous herein
It states.
Stage two: network model training may include following steps S704-S705.
S704, by the positive sample feature of at least one above-mentioned each sample user characteristics pair of sample of users feature centering and negative sample
Eigen is as the input for applying recommendation network model.
S705 passes through the above-mentioned positive sample feature using recommendation network model to above-mentioned each sample user characteristics pair and negative sample
Eigen is learnt.
In some possible embodiments, above-mentioned to be carried out by least one above-mentioned sample of users feature to network model
Trained realization can be found in the implementation that step S404 and S405 are provided in above-mentioned embodiment shown in Fig. 4, no longer superfluous herein
It states.
Stage three: network model test may include following steps S706.
S706 is obtained and is recommended the test feature of test for application and input using recommendation network model.
In some possible embodiments, the realization of the above-mentioned training by above-mentioned network model can be found in above-mentioned Fig. 4 institute
The implementation that step S46 is provided in the embodiment shown, details are not described herein.
In some possible embodiments, terminal by it is above-mentioned using recommendation network model to above-mentioned application to be recommended
Second attributive character is learnt, and when determining the corresponding user's recommendation of above-mentioned application to be recommended, to be recommended can also be answered above-mentioned
Second user attributive character and above-mentioned first application attribute feature input above-mentioned application recommendation network model, are answered by above-mentioned
Above-mentioned second user attributive character is learnt with recommendation network model, and is exported corresponding to above-mentioned second user attributive character
Above-mentioned second user recommend it is above-mentioned first application user's recommendation.
In one possible implementation, any of the above-described sample of users corresponding to positive sample user in i first is default pushes away
Recommend value, the corresponding second default recommendation of negative sample user.Terminal can be by any sample of users feature to i10Positive sample feature i1
With negative sample feature i0Above-mentioned application recommendation network model is inputted, and obtains the above-mentioned positive sample exported using recommendation network model
Feature i1Corresponding first user recommendation and negative sample feature i0Corresponding second user recommendation.Terminal can be according to above-mentioned
The difference of one user's recommendation and above-mentioned second user recommendation is preset in conjunction with the above-mentioned first default recommendation and above-mentioned second and is pushed away
The difference for recommending value calculates recommendation loss, corrects above-mentioned application recommendation network model according to the loss of above-mentioned recommendation, adjusts above-mentioned
Using recommendation network model to the precision of prediction of the corresponding user's recommendation of any user.
In some possible embodiments, terminal, which orients the first application to second user, utilizes front training when recommending
Good predicts that this is using recommendation network model to the user group interested of the first application using recommendation network model
Output score (corresponding to user's recommendation) represents the interest level to the first interested second user of application, Jin Erke
The recommendation of second user is oriented to the first application according to user's recommendation.
In the embodiment of the present application, the application attribute feature of the first application can be used for the corresponding application of second user by terminal
In the building of feature, and then it can will be obtained constructed by the user property feature of the application attribute feature of the first application and second user
Application feature be sent into and the user characteristics of second user carried out using in recommendation network model, passing through application recommendation network model
Study can be realized with second user corresponding recommendation when the first application of recommendation of pre- direction finding second user to second user
Orientation recommends the first application, easy to operate, enhances using recommendation and the affinity that is associated with of user, and then can reduce and push away to user
The probability of the game of non-user demand or hobby is recommended, improves the accuracy rate that application is recommended, while can also reduce using recommendation
Redundancy rate, the user's viscosity for enhancing terminal.
Description based on the embodiment of the method that above-mentioned application is recommended, the embodiment of the present application also disclose what a kind of application was recommended
Device (may be simply referred to as convenience of description using recommendation apparatus), this can be operate in a meter in terminal using recommendation apparatus
Calculation machine program (including program code), this can be applied to the application of Fig. 2-embodiment illustrated in fig. 7 using recommendation apparatus and recommend
In method, for executing the step applied in recommended method.For convenience of description, will be illustrated by taking terminal as an example below.Please
It is a structural schematic diagram of terminal provided by the embodiments of the present application referring to Fig. 8, Fig. 8.In the embodiment of the present application, which can transport
Row such as lower unit:
Feature acquiring unit 81, for obtaining the application feature of application to be recommended.
It wherein, include the first attributive character and the second attributive character in the application feature of above-mentioned application to be recommended, above-mentioned the
One attributive character is the first user property feature, above-mentioned second attributive character is the second application attribute feature or above-mentioned first
Attributive character is the first application attribute feature, above-mentioned second attributive character is second user attributive character.
Characteristic processing unit 82, first of the above-mentioned application to be recommended for being obtained according to features described above acquiring unit 81 belong to
Property feature determine and apply recommendation network model, and by it is above-mentioned using recommendation network model to the second of above-mentioned application to be recommended the category
Property feature learnt, determine the corresponding application recommendation of above-mentioned application to be recommended or above-mentioned to be recommended apply corresponding user
Recommendation.
Wherein, above-mentioned to be obtained using recommendation network model by the sample application feature training of the first user-association, Huo Zheshang
It states and is obtained using recommendation network model by the sample of users feature training of the first association.
Recommend predicting unit 83, the above-mentioned application recommendation for determining according to features described above processing unit 82 determines upward
State the application priority that the first user recommends above-mentioned application to be recommended, or the above-mentioned use determined according to features described above processing unit
Family recommendation determines the User Priority for recommending above-mentioned first application.
In a kind of feasible implementation, above-mentioned recommendation predicting unit 83 is also used to:
When the application priority of above-mentioned application to be recommended be more than or equal to it is default using priority threshold value when, Xiang Shangshu the
One user recommends the application to be recommended;Or
When the User Priority of above-mentioned application to be recommended is more than or equal to pre-set user priority threshold value, determine to the
Two users recommend above-mentioned first application.
In a kind of feasible implementation, above-mentioned first attributive character is the first user property feature, above-mentioned second category
Property feature be the second application attribute feature;
Features described above processing unit 82 is used for:
It will be wrapped in the first user property feature that features described above acquiring unit 81 obtains and application recommendation network model set
Include it is each matched using the user property feature of recommendation network model interaction, from above-mentioned application recommendation network model set
It determines to apply recommendation network model associated by corresponding above-mentioned first user of above-mentioned first user property feature;
Wherein, above-mentioned using the user in recommendation network model set further including other users except above-mentioned first user
Other application recommendation network model associated by attributive character, above-mentioned other application recommendation network model are closed by above-mentioned other users
The sample application feature training of connection obtains.
In a kind of feasible implementation, features described above acquiring unit 81 is also used to:
Obtain the sample data for recommending at least two samples application of training for application, wherein any sample application
It include above-mentioned first user attribute data and sample application attribute data in sample data;
At least one sample application feature pair is constructed according to the sample data that above-mentioned at least two sample is applied, wherein one
A sample application feature centering includes a positive sample feature and a negative sample feature, wherein is wrapped in above-mentioned positive sample feature
Above-mentioned first user property feature and positive sample application attribute feature are included, includes that above-mentioned first user belongs in above-mentioned negative sample feature
Property feature and negative sample application attribute feature;
Recommendation network model is applied to building according at least one above-mentioned sample application feature.
It include active degree in the sample application attribute data of above-mentioned various kinds this application in a kind of feasible implementation
Indicate information;
Features described above acquiring unit 81 is used for:
The application of above-mentioned at least two sample is matched two-by-two, obtains the application pair of at least one sample;
At least one above-mentioned any sample of sample application centering is applied, i is performed the following operations to obtain at least one sample
Using feature pair:
According to above-mentioned sample apply to the active degree instruction information of two samples of i application determine positive sample apply and
Negative sample application, wherein the active degree of above-mentioned positive sample application is higher than the active degree of above-mentioned negative sample application;
Positive sample is constructed according to the sample application attribute data that above-mentioned first user attribute data and above-mentioned positive sample are applied
Feature i1, and negative sample is constructed according to the sample application attribute data that above-mentioned first user attribute data and above-mentioned negative sample are applied
Feature i0, obtain above-mentioned sample and apply to the corresponding sample application feature of i to i10;
Wherein, above-mentioned sample application feature is to i10In include above-mentioned positive sample feature i1With above-mentioned negative sample feature i0。
In a kind of feasible implementation, features described above acquiring unit 81 is used for:
The positive sample feature of at least one above-mentioned sample application feature centering various kinds this application feature pair and negative sample is special
Sign as using recommendation network model input, by it is above-mentioned using recommendation network model to above-mentioned various kinds this application feature pair
Positive sample feature and negative sample feature are learnt, any using the corresponding ability using recommendation of feature to obtain prediction;
Wherein, the corresponding application recommendation of any sample application feature centering positive sample feature is corresponding greater than negative sample feature
Application recommendation.
In a kind of feasible implementation, features described above processing unit 82 is used for:
Second application attribute feature of above-mentioned application to be recommended and above-mentioned first user property feature are inputted into above-mentioned application
Recommendation network model learns above-mentioned second application attribute feature by above-mentioned application recommendation network model, and export to
Corresponding above-mentioned first user of above-mentioned first user property feature recommends the application recommendation of above-mentioned application to be recommended.
In a kind of feasible implementation, any of the above-described sample is applied to be pushed away to positive sample application corresponding first is default in i
Recommend value, the corresponding second default recommendation of negative sample application;
Features described above acquiring unit 81 is also used to:
Obtain the above-mentioned positive sample feature i exported using recommendation network model1Corresponding first applies recommendation and negative sample
Eigen i0Corresponding second applies recommendation;
According to the difference of above-mentioned first application recommendation and above-mentioned second application recommendation, in conjunction with the above-mentioned first default recommendation
The difference of value and the above-mentioned second default recommendation calculates recommendation loss;
Above-mentioned application recommendation network model is corrected according to the loss of above-mentioned recommendation, adjusts above-mentioned application recommendation network model pair
It is any using the corresponding precision of prediction using recommendation.
In a kind of feasible implementation, above-mentioned first attributive character is the first application attribute feature, above-mentioned second category
Property feature be second user attributive character;
Features described above processing unit 81 is used for:
Each application recommendation network mould that will include in above-mentioned first application attribute feature and application recommendation network model set
The associated application attribute feature of type is matched, and determines that above-mentioned first application belongs to from above-mentioned application recommendation network model set
Property corresponding above-mentioned first application of feature is associated applies recommendation network model;
Wherein, it is above-mentioned using further include in recommendation network model set it is above-mentioned first application except other application application
Other application recommendation network model associated by attributive character, above-mentioned other application recommendation network model are closed by above-mentioned other application
The sample of users feature training of connection obtains.
In a kind of feasible implementation, features described above acquiring unit 81 is also used to:
Obtain the user data for recommending at least two sample of users of training for application, wherein any sample of users
It include above-mentioned first application attribute data and sample of users attribute data in user data;
At least one sample of users feature pair is constructed according to the user data of above-mentioned at least two sample of users, according to above-mentioned
At least one sample of users feature applies recommendation network model to building, wherein a sample of users feature centering includes one
Positive sample feature and a negative sample feature, wherein including above-mentioned first application attribute feature and just in above-mentioned positive sample feature
Sample of users attributive character includes that above-mentioned first application attribute feature and negative sample user property are special in above-mentioned negative sample feature
Sign.
It include active degree in the sample of users attribute data of above-mentioned each sample of users in a kind of feasible implementation
Indicate information;
Features described above acquiring unit 81 is used for:
Above-mentioned at least two sample of users is matched two-by-two, obtains at least one sample of users pair;
At least one above-mentioned any sample of users of sample of users centering performs the following operations i to obtain at least one sample
User characteristics pair:
According to above-mentioned sample of users to the active degrees of two sample of users of i instruction information determine positive sample user and
Negative sample user, wherein the active degree of above-mentioned positive sample user is higher than the active degree of above-mentioned negative sample user;
Positive sample is constructed according to the sample of users attribute data of above-mentioned first application attribute data and above-mentioned positive sample user
Feature i1, and negative sample is constructed according to the sample of users attribute data of above-mentioned first application attribute data and above-mentioned negative sample user
Feature i0, above-mentioned sample of users is obtained to the corresponding sample of users feature of i to i10;
Wherein, above-mentioned sample of users feature is to i10In include above-mentioned positive sample feature i1With above-mentioned negative sample feature i0。
In a kind of feasible implementation, features described above acquiring unit 81 is used for:
The positive sample feature of at least one above-mentioned each sample user characteristics pair of sample of users feature centering and negative sample is special
Sign as using recommendation network model input, by it is above-mentioned using recommendation network model to above-mentioned each sample user characteristics pair
Positive sample feature and negative sample feature are learnt, and predict any ability using the corresponding user's recommendation of feature to obtain;
Wherein, the corresponding user's recommendation of any sample of users feature centering positive sample feature is corresponding greater than negative sample feature
User's recommendation.
In a kind of feasible implementation, features described above processing unit 82 is used for:
The second user attributive character of above-mentioned application to be recommended and above-mentioned first application attribute feature are inputted into above-mentioned application
Recommendation network model learns above-mentioned second user attributive character by above-mentioned application recommendation network model, and export to
The corresponding above-mentioned second user of above-mentioned second user attributive character recommends user's recommendation of above-mentioned first application.
In a kind of feasible implementation, any of the above-described sample of users corresponding to positive sample user in i first is default to be pushed away
Recommend value, the corresponding second default recommendation of negative sample user;
Features described above processing unit 82 is also used to:
Obtain the above-mentioned positive sample feature i exported using recommendation network model1Corresponding first user recommendation and negative sample
Eigen i0Corresponding second user recommendation;
According to the difference of above-mentioned first user recommendation and above-mentioned second user recommendation, in conjunction with the above-mentioned first default recommendation
The difference of value and the above-mentioned second default recommendation calculates recommendation loss;
Above-mentioned application recommendation network model is corrected according to the loss of above-mentioned recommendation, adjusts above-mentioned application recommendation network model pair
The precision of prediction of the corresponding user's recommendation of any user.
In a kind of feasible implementation, the user property feature of above-mentioned first user and/or second user is by user
Age, user's gender, user's educational background, region and user locating for user are true using any user attribute data in account
It is fixed;
The sample application attribute feature of application attribute feature and/or the sample application of above-mentioned application to be recommended is marked by application
At least one of knowledge, application type, active degree instruction information, application resource type and user behavior data application attribute
Data determine.
According to an embodiment of the present application one, in above-mentioned application recommended method shown in Fig. 2 described by step S201 to S203
Implementation can each unit of terminal as shown in Figure 8 execute.For example, being walked in above-mentioned application recommended method shown in Fig. 2
Implementation described in rapid S201, S202 and S203 can distinguish feature acquiring unit 81 in terminal as shown in Figure 8, at feature
Reason unit 82 is executed with predicting unit 83 is recommended.Wherein, features described above acquiring unit 81, characteristic processing unit 82 and recommendation
Implementation performed by predicting unit 83 can be found in implementation provided by each step in above-described embodiment one, herein not
It repeats again.
According to an embodiment of the present application two, in above-mentioned application recommended method shown in Fig. 4 described by step S401 to S406
Implementation terminal also as shown in Figure 8 in feature acquiring unit 81 and characteristic processing unit 82 execute.Wherein, above-mentioned spy
Implementation performed by sign acquiring unit 81 and characteristic processing unit 82 can be found in each step in above-described embodiment two and be retouched
The implementation stated, details are not described herein.
According to an embodiment of the present application three, in above-mentioned application recommended method shown in fig. 6 described by step S601 to S604
Implementation each unit can execute in terminal as shown in Figure 8.For example, being walked in above-mentioned application recommended method shown in fig. 6
Implementation described in rapid S601-S602, S603 and S604 can feature acquiring unit 81 in terminal respectively as shown in Figure 8,
Characteristic processing unit 82 is executed with predicting unit 83 is recommended.Wherein, features described above acquiring unit 81, characteristic processing unit 82 with
And implementation performed by predicting unit 83 is recommended to can be found in implementation provided by each step in above-described embodiment three,
Details are not described herein.
According to an embodiment of the present application four, in above-mentioned application recommended method shown in Fig. 7 described by step S701 to S706
Implementation terminal also as shown in Figure 8 in feature acquiring unit 81 and characteristic processing unit 82 execute.Wherein, above-mentioned spy
Implementation performed by sign acquiring unit 81 and characteristic processing unit 82 can be found in each step in above-described embodiment four and be retouched
The implementation stated, details are not described herein.
In the embodiment of the present application, each unit in above-mentioned terminal shown in Fig. 8 can respectively or all merge into one
A or several other units come constitute or some (a little) unit therein can also be split as again it is functionally smaller more
A unit is constituted, this may be implemented similarly to operate, the realization of the technical effect without influencing the embodiment of the present application.Above-mentioned list
Member is logic-based function division, and in practical applications, the function of a unit can also be realized by multiple units, or
The function of multiple units is realized by a unit.In other feasible implementations of the application, above-mentioned terminal also be can wrap
Other units are included, in practical applications, these functions can also be assisted to realize by other units, and can be assisted by multiple units
It realizes, herein with no restrictions.
In the embodiment of the present application, the user property feature of the first user can be used for the application feature of the second application by terminal
Building in, and then can by the user property feature of the first user and second apply application attribute feature constructed by obtain answer
It is sent into feature using in recommendation network model, is learnt by application feature of the application recommendation network model to the second application
Recommend recommendation when the second application with pre- the first user of direction finding, and then can realize the orientation recommendation to the first user, operation letter
It is single.Further, in the embodiment of the present application, the user property feature of second user can be used for the application of the first application by terminal
In the building of feature, and then it can will be obtained constructed by the user property feature of second user and the application attribute feature of the first application
Application feature be sent into and the user characteristics of second user carried out using in recommendation network model, passing through application recommendation network model
User's recommendation when the first application is recommended in study with pre- direction finding second user, and then can be realized and be recommended first using orientation
Second user.The user property feature of first user is dissolved into the recommendation process of the second application by the embodiment of the present application, and/or
The application attribute feature of first application is dissolved into the application recommendation process of the first user, is enhanced using recommendation with user's
It is associated with affinity, and then the probability for recommending the game of non-user demand or hobby to user can be reduced, improves what application was recommended
Accuracy rate, while the user's viscosity for enhancing terminal using the redundancy rate recommended can also be reduced.
Recommended method is applied shown in based on the above embodiment, the embodiment of the present application also provides a kind of terminals, which can
To be applied in Fig. 2-embodiment illustrated in fig. 7 application recommended method, for executing the step applied in recommended method.
It is another structural schematic diagram of terminal provided by the embodiments of the present application referring to Fig. 9, Fig. 9.In the embodiment of the present application,
Above-mentioned terminal may include processor 91, computer storage medium (or memory) 92 and communication interface 93.Wherein, above-mentioned calculating
Machine storage medium (or memory) 92 is used to store the computer journey using recommended method for supporting the various embodiments described above to provide
Sequence, above-mentioned computer program can be one or more program instruction (or referred to as instruct, such as instruction 1, instruction 2 ... instruction
N).Wherein, above-mentioned processor 91, computer storage medium (or memory) 92 and communication interface 93 can by bus 94 or its
He connects mode, in Fig. 9 shown in the embodiment of the present application for being connected by bus 94.
Communication interface is to realize the medium interacted between terminal and external equipment with information exchange.Processor (such as
Central processing unit (central processing unit, CPU)) be terminal calculating core and control core, be suitable for real
Existing one or one or more instruction are particularly adapted to load and execute one or one or more instruct to realize correlation method process
Or corresponding function;Processor provided by the embodiments of the present application is for obtaining using feature, handling using feature and determine application
Recommendation, etc..Computer storage medium (Memory) is the memory device in server, for storing program and data.It can
With understanding, computer storage medium herein both may include the built-in storage medium of terminal, naturally it is also possible to including end
Hold supported expansion storage medium.Computer storage medium provides memory space, which stores the operation of terminal
System.Also, it also houses and is suitable for by one or more than one instructions that processor loads and executes in the memory space,
These instructions can be one or more computer program (including program code).It should be noted that meter herein
Calculation machine storage medium can be high speed RAM memory, be also possible to non-labile memory (non-volatile
Memory), a for example, at least magnetic disk storage;It optionally can also be that at least one is located remotely from the calculating of aforementioned processor
Machine storage medium.
In the embodiment of the present application, processor load and execute one stored in computer storage medium or one or more
Instruction, to realize corresponding steps of the above-mentioned Fig. 2 into Fig. 7 in method flow provided by each embodiment.In the specific implementation, meter
One in calculation machine storage medium or one or more instruction are loaded by processor and execute following steps:
Obtain the application feature of application to be recommended, wherein include the first attribute in the application feature of above-mentioned application to be recommended
Feature and the second attributive character, above-mentioned first attributive character is the first user property feature, above-mentioned second attributive character is second
Application attribute feature or above-mentioned first attributive character are the first application attribute feature, above-mentioned second attributive character is the second use
Family attributive character;
It is determined according to the first attributive character of above-mentioned application to be recommended and applies recommendation network model, and pushed away by above-mentioned application
It recommends network model to learn the second attributive character of above-mentioned application to be recommended, determines the corresponding application of above-mentioned application to be recommended
Recommendation or the corresponding user's recommendation of above-mentioned application to be recommended, wherein above-mentioned application recommendation network model is by the first user
Associated sample application feature training obtains or the above-mentioned sample of users spy using recommendation network model by the first association
Sign training obtains;
Recommend above-mentioned application to be recommended according to determining using recommendation to above-mentioned first user for above-mentioned application to be recommended
Using priority, or is determined according to user's recommendation of above-mentioned application to be recommended and recommend the user of above-mentioned first application preferential
Grade.
In some possible embodiments, one in above-mentioned processor load computer storage medium or one or more
Instruction execution following steps:
When the application priority of above-mentioned application to be recommended be more than or equal to it is default using priority threshold value when, Xiang Shangshu the
One user recommends above-mentioned application to be recommended;Or
When the User Priority of above-mentioned application to be recommended is more than or equal to pre-set user priority threshold value, determine to the
Two users recommend above-mentioned first application.
In some possible embodiments, above-mentioned first attributive character is the first user property feature, above-mentioned second category
Property feature be the second application attribute feature;One in above-mentioned processor load computer storage medium or one or more instruction are held
The provided implementation of the step of row is determined according to the first attributive character of above-mentioned application to be recommended using recommendation network model,
It is specific to execute following steps:
Each application recommendation network mould that will include in above-mentioned first user property feature and application recommendation network model set
The associated user property feature of type is matched, and determines that above-mentioned first user belongs to from above-mentioned application recommendation network model set
Property corresponding above-mentioned first user of feature associated by apply recommendation network model;
Wherein, above-mentioned using the user in recommendation network model set further including other users except above-mentioned first user
Other application recommendation network model associated by attributive character, above-mentioned other application recommendation network model are closed by above-mentioned other users
The sample application feature training of connection obtains.
In some possible embodiments, one in above-mentioned processor load computer storage medium or one or more
Instruction execution following steps:
Obtain the sample data for recommending at least two samples application of training for application, wherein any sample application
It include above-mentioned first user attribute data and sample application attribute data in sample data;
At least one sample application feature pair is constructed according to the sample data that above-mentioned at least two sample is applied, wherein one
A sample application feature centering includes a positive sample feature and a negative sample feature, wherein is wrapped in above-mentioned positive sample feature
Above-mentioned first user property feature and positive sample application attribute feature are included, includes that above-mentioned first user belongs in above-mentioned negative sample feature
Property feature and negative sample application attribute feature;
Recommendation network model is applied to building according at least one above-mentioned sample application feature.
It in some possible embodiments, include active degree in the sample application attribute data of above-mentioned various kinds this application
Indicate information;One in above-mentioned processor load computer storage medium or one or more instruction execution are according to above-mentioned at least two
The sample data of a sample application constructs implementation provided by the step of at least one sample application feature pair, specific to execute
Following steps:
The application of above-mentioned at least two sample is matched two-by-two, obtains the application pair of at least one sample;
At least one above-mentioned any sample of sample application centering is applied, i is performed the following operations to obtain at least one sample
Using feature pair:
According to above-mentioned sample apply to the active degree instruction information of two samples of i application determine positive sample apply and
Negative sample application, wherein the active degree of above-mentioned positive sample application is higher than the active degree of above-mentioned negative sample application;
Positive sample is constructed according to the sample application attribute data that above-mentioned first user attribute data and above-mentioned positive sample are applied
Feature i1, and negative sample is constructed according to the sample application attribute data that above-mentioned first user attribute data and above-mentioned negative sample are applied
Feature i0, obtain above-mentioned sample and apply to the corresponding sample application feature of i to i10;
Wherein, above-mentioned sample application feature is to i10In include above-mentioned positive sample feature i1With above-mentioned negative sample feature i0。
In some possible embodiments, one in above-mentioned processor load computer storage medium or one or more
The provided realization of the step of instruction execution applies recommendation network model to building according at least one above-mentioned sample application feature
Mode specifically executes following steps:
The positive sample feature of at least one above-mentioned sample application feature centering various kinds this application feature pair and negative sample is special
Sign as using recommendation network model input, by it is above-mentioned using recommendation network model to above-mentioned various kinds this application feature pair
Positive sample feature and negative sample feature are learnt, any using the corresponding ability using recommendation of feature to obtain prediction;
Wherein, the corresponding application recommendation of any sample application feature centering positive sample feature is corresponding greater than negative sample feature
Application recommendation.
In some possible embodiments, one in above-mentioned processor load computer storage medium or one or more
Instruction execution is learnt by above-mentioned using second attributive character of the recommendation network model to above-mentioned application to be recommended, in determination
The provided implementation of the step of stating application to be recommended corresponding application recommendation, specifically executes following steps:
Second application attribute feature of above-mentioned application to be recommended and above-mentioned first user property feature are inputted into above-mentioned application
Recommendation network model learns above-mentioned second application attribute feature by above-mentioned application recommendation network model, and export to
Corresponding above-mentioned first user of above-mentioned first user property feature recommends the application recommendation of above-mentioned application to be recommended.
In some possible embodiments, any of the above-described sample is applied pushes away to positive sample application corresponding first is default in i
Recommend value, the corresponding second default recommendation of negative sample application;By any sample application feature to i10Positive sample feature i1With it is negative
Sample characteristics i0Input after above-mentioned application recommendation network model, one in above-mentioned processor load computer storage medium or
One or more instruction execution following steps:
Obtain the above-mentioned positive sample feature i exported using recommendation network model1Corresponding first applies recommendation and negative sample
Eigen i0Corresponding second applies recommendation;
According to the difference of above-mentioned first application recommendation and above-mentioned second application recommendation, in conjunction with the above-mentioned first default recommendation
The difference of value and the above-mentioned second default recommendation calculates recommendation loss;
Above-mentioned application recommendation network model is corrected according to the loss of above-mentioned recommendation, adjusts above-mentioned application recommendation network model pair
It is any using the corresponding precision of prediction using recommendation.
In some possible embodiments, above-mentioned first attributive character is the first application attribute feature, above-mentioned second category
Property feature be second user attributive character;One in above-mentioned processor load computer storage medium or one or more instruction are held
The provided implementation of the step of row is determined according to the first attributive character of above-mentioned application to be recommended using recommendation network model,
It is specific to execute following steps:
Each application recommendation network mould that will include in above-mentioned first application attribute feature and application recommendation network model set
The associated application attribute feature of type is matched, and determines that above-mentioned first application belongs to from above-mentioned application recommendation network model set
Property corresponding above-mentioned first application of feature is associated applies recommendation network model;
Wherein, it is above-mentioned using further include in recommendation network model set it is above-mentioned first application except other application application
Other application recommendation network model associated by attributive character, above-mentioned other application recommendation network model are closed by above-mentioned other application
The sample of users feature training of connection obtains.
In some possible embodiments, one in above-mentioned processor load computer storage medium or one or more
Instruction execution following steps:
Obtain the user data for recommending at least two sample of users of training for application, wherein any sample of users
It include above-mentioned first application attribute data and sample of users attribute data in user data;
At least one sample of users feature pair is constructed according to the user data of above-mentioned at least two sample of users, according to above-mentioned
At least one sample of users feature applies recommendation network model to building, wherein a sample of users feature centering includes one
Positive sample feature and a negative sample feature, wherein including above-mentioned first application attribute feature and just in above-mentioned positive sample feature
Sample of users attributive character includes that above-mentioned first application attribute feature and negative sample user property are special in above-mentioned negative sample feature
Sign.
It in some possible embodiments, include active degree in the sample of users attribute data of above-mentioned each sample of users
Indicate information;One in above-mentioned processor load computer storage medium or one or more instruction execution are according to above-mentioned at least two
The user data of a sample of users constructs implementation provided by the step of at least one sample of users feature pair, specific to execute
Following steps:
Above-mentioned at least two sample of users is matched two-by-two, obtains at least one sample of users pair;
At least one above-mentioned any sample of users of sample of users centering performs the following operations i to obtain at least one sample
User characteristics pair:
According to above-mentioned sample of users to the active degrees of two sample of users of i instruction information determine positive sample user and
Negative sample user, wherein the active degree of above-mentioned positive sample user is higher than the active degree of above-mentioned negative sample user;
Positive sample is constructed according to the sample of users attribute data of above-mentioned first application attribute data and above-mentioned positive sample user
Feature i1, and negative sample is constructed according to the sample of users attribute data of above-mentioned first application attribute data and above-mentioned negative sample user
Feature i0, above-mentioned sample of users is obtained to the corresponding sample of users feature of i to i10;
Wherein, above-mentioned sample of users feature is to i10In include above-mentioned positive sample feature i1With above-mentioned negative sample feature i0。
In some possible embodiments, one in above-mentioned processor load computer storage medium or one or more
The provided realization of the step of instruction execution applies recommendation network model to building according at least one above-mentioned sample of users feature
Mode specifically executes following steps:
The positive sample feature of at least one above-mentioned each sample user characteristics pair of sample of users feature centering and negative sample is special
Sign as using recommendation network model input, by it is above-mentioned using recommendation network model to above-mentioned each sample user characteristics pair
Positive sample feature and negative sample feature are learnt, and predict any ability using the corresponding user's recommendation of feature to obtain;
Wherein, the corresponding user's recommendation of any sample of users feature centering positive sample feature is corresponding greater than negative sample feature
User's recommendation.
In some possible embodiments, one in above-mentioned processor load computer storage medium or one or more
Instruction execution is learnt by above-mentioned using second attributive character of the recommendation network model to above-mentioned application to be recommended, in determination
The provided implementation of the step of stating application to be recommended corresponding user's recommendation, specifically executes following steps:
The second user attributive character of above-mentioned application to be recommended and above-mentioned first application attribute feature are inputted into above-mentioned application
Recommendation network model learns above-mentioned second user attributive character by above-mentioned application recommendation network model, and export to
The corresponding above-mentioned second user of above-mentioned second user attributive character recommends user's recommendation of above-mentioned first application.
In some possible embodiments, any of the above-described sample of users corresponding to positive sample user in i first is default pushes away
Recommend value, the corresponding second default recommendation of negative sample user;
By any sample of users feature to i10Positive sample feature i1With negative sample feature i0Above-mentioned application is inputted to recommend
After network model, one or one or more instruction execution following steps in above-mentioned processor load computer storage medium:
Obtain the above-mentioned positive sample feature i exported using recommendation network model1Corresponding first user recommendation and negative sample
Eigen i0Corresponding second user recommendation;
According to the difference of above-mentioned first user recommendation and above-mentioned second user recommendation, in conjunction with the above-mentioned first default recommendation
The difference of value and the above-mentioned second default recommendation calculates recommendation loss;
Above-mentioned application recommendation network model is corrected according to the loss of above-mentioned recommendation, adjusts above-mentioned application recommendation network model pair
The precision of prediction of the corresponding user's recommendation of any user.
In some possible embodiments, the user property feature of above-mentioned first user and/or above-mentioned second user by
Age of user, user's gender, user's educational background, region and user locating for user are using any user attribute data in account
It determines;
The sample application attribute feature of application attribute feature and/or the sample application of above-mentioned application to be recommended is marked by application
At least one of knowledge, application type, active degree instruction information, application resource type and user behavior data application attribute
Data determine.
In the embodiment of the present application, the user property feature of the first user can be used for the application feature of the second application by terminal
Building in, and then can by the user property feature of the first user and second apply application attribute feature constructed by obtain answer
It is sent into feature using in recommendation network model, is learnt by application feature of the application recommendation network model to the second application
Recommend recommendation when the second application with pre- the first user of direction finding, and then can realize the orientation recommendation to the first user, operation letter
It is single.Further, in the embodiment of the present application, the user property feature of second user can be used for the application of the first application by terminal
In the building of feature, and then it can will be obtained constructed by the user property feature of second user and the application attribute feature of the first application
Application feature be sent into and the user characteristics of second user carried out using in recommendation network model, passing through application recommendation network model
User's recommendation when the first application is recommended in study with pre- direction finding second user, and then can be realized and be recommended first using orientation
Second user.The user property feature of first user is dissolved into the recommendation process of the second application by the embodiment of the present application, and/or
The application attribute feature of first application is dissolved into the application recommendation process of the first user, is enhanced using recommendation with user's
It is associated with affinity, and then the probability for recommending the game of non-user demand or hobby to user can be reduced, improves what application was recommended
Accuracy rate, while the user's viscosity for enhancing terminal using the redundancy rate recommended can also be reduced.
Those of ordinary skill in the art will appreciate that realizing all or part of the process in above-described embodiment method, being can be with
Relevant hardware is instructed to complete by computer program, the program can be stored in a computer-readable storage medium
In, the program is when being executed, it may include such as the process of the embodiment of above-mentioned each method.Wherein, the storage medium can be magnetic
Dish, CD, read-only memory (Read-Only Memory, ROM) or random access memory (Random Access
Memory, RAM) etc..
Claims (15)
1. a kind of method that application is recommended, which is characterized in that the described method includes:
Obtain the application feature of application to be recommended, wherein include the first attributive character in the application feature of the application to be recommended
With the second attributive character, first attributive character is the first user property feature, second attributive character is the second application
Attributive character or first attributive character are the first application attribute feature, second attributive character is second user category
Property feature;
It is determined according to the first attributive character of the application to be recommended and applies recommendation network model, and net is recommended by the application
Network model learns the second attributive character of the application to be recommended, determines that the corresponding application of the application to be recommended is recommended
Value or the corresponding user's recommendation of the application to be recommended, wherein the application recommendation network model is by the first user-association
Sample application feature training obtain or it is described using recommendation network model by the first association sample of users feature instruct
It gets;
According to the application to be recommended using the determining application for recommending the application to be recommended to first user of recommendation
Priority, or the User Priority for recommending first application is determined according to user's recommendation of the application to be recommended.
2. the method according to claim 1, wherein the method also includes:
When the application priority of the application to be recommended, which is more than or equal to, to be preset using priority threshold value, Xiang Suoshu first is used
Recommend the application to be recommended in family;Or
When the User Priority of the application to be recommended is more than or equal to pre-set user priority threshold value, determines and used to second
Recommend first application in family.
3. method according to claim 1 or 2, which is characterized in that first attributive character is that the first user property is special
Sign, second attributive character are the second application attribute feature;
It is described to include: using recommendation network model according to the determination of the first attributive character of the application to be recommended
Each application recommendation network model for including in the first user property feature and application recommendation network model set is closed
The user property feature of connection is matched, and determines that first user property is special from the application recommendation network model set
It levies and applies recommendation network model associated by corresponding first user;
Wherein, described using the user property in recommendation network model set further including other users except first user
Other application recommendation network model associated by feature, the other application recommendation network model are associated by the other users
The training of sample application feature obtains.
4. according to the method described in claim 3, it is characterized in that, before the application feature for obtaining application to be recommended, institute
State method further include:
Obtain the sample data for recommending at least two samples application of training for application, wherein the sample of any sample application
It include first user attribute data and sample application attribute data in data;
At least one sample application feature pair is constructed according to the sample data that at least two sample is applied, wherein a sample
This application feature centering includes a positive sample feature and a negative sample feature, wherein includes institute in the positive sample feature
The first user property feature and positive sample application attribute feature are stated, includes that first user property is special in the negative sample feature
It seeks peace negative sample application attribute feature;
Recommendation network model is applied to building according at least one described sample application feature.
5. according to the method described in claim 4, it is characterized in that, it is described by the application recommendation network model to it is described to
Recommend the second attributive character of application to be learnt, determines that the corresponding application recommendation of the application to be recommended includes:
Second application attribute feature of the application to be recommended and the first user property feature are inputted the application to recommend
Network model learns the second application attribute feature by the application recommendation network model, and exports to described
Corresponding first user of first user property feature recommends the application recommendation of the application to be recommended.
6. method according to claim 1 or 2, which is characterized in that first attributive character is that the first application attribute is special
Sign, second attributive character are second user attributive character;
It is described to include: using recommendation network model according to the determination of the first attributive character of the application to be recommended
Each application recommendation network model for including in the first application attribute feature and application recommendation network model set is closed
The application attribute feature of connection is matched, and determines that first application attribute is special from the application recommendation network model set
It levies corresponding described first and applies recommendation network model using associated;
Wherein, it is described using further include in recommendation network model set it is described first application except other application application attribute
Other application recommendation network model associated by feature, the other application recommendation network model are associated by the other application
The training of sample of users feature obtains.
7. according to the method described in claim 6, it is characterized in that, before the application feature for obtaining application to be recommended, institute
State method further include:
Obtain the user data for recommending at least two sample of users of training for application, wherein the user of any sample of users
It include the first application attribute data and sample of users attribute data in data;
Construct at least one sample of users feature pair according to the user data of at least two sample of users, according to it is described at least
One sample of users feature applies recommendation network model to building, wherein a sample of users feature centering includes a positive sample
Eigen and a negative sample feature, wherein include the first application attribute feature and positive sample in the positive sample feature
User property feature includes the first application attribute feature and negative sample user property feature in the negative sample feature.
8. the method according to the description of claim 7 is characterized in that it is described by the application recommendation network model to it is described to
Recommend the second attributive character of application to be learnt, determines that the corresponding user's recommendation of the application to be recommended includes:
The second user attributive character of the application to be recommended and the first application attribute feature are inputted the application to recommend
Network model learns the second user attributive character by the application recommendation network model, and exports to described
The corresponding second user of second user attributive character recommends user's recommendation of first application.
9. according to the described in any item methods of claim 3-8, which is characterized in that first user and/or second use
The user property feature at family region as locating for age of user, user's gender, user's educational background, user and user are using in account
Any user attribute data determines;
The application to be recommended application attribute feature and/or sample application sample application attribute feature by application identities, answer
With at least one of type, active degree instruction information, application resource type and user behavior data application attribute data
It determines.
10. a kind of terminal, which is characterized in that the terminal includes:
Feature acquiring unit, for obtaining the application feature of application to be recommended, wherein in the application feature of the application to be recommended
Including the first attributive character and the second attributive character, first attributive character is the first user property feature, second category
Property feature is the second application attribute feature or first attributive character is the first application attribute feature, second attribute
Feature is second user attributive character;
First attributive character of characteristic processing unit, the application to be recommended for being obtained according to the feature acquiring unit is true
Surely apply recommendation network model, and by it is described using recommendation network model to the second attributive character of the application to be recommended into
Row study determines the corresponding application recommendation of the application to be recommended or the corresponding user's recommendation of the application to be recommended,
Wherein, described to be obtained or the application pushes away using recommendation network model by the sample application feature training of the first user-association
Network model is recommended to be obtained by the sample of users feature training of the first association;
Recommend predicting unit, the application recommendation for determining according to the characteristic processing unit is determined to be used to described first
Recommend the application priority of the application to be recommended, or the user's recommendation determined according to the characteristic processing unit in family
Determine the User Priority for recommending first application.
11. terminal according to claim 10, which is characterized in that the recommendation predicting unit is also used to:
When the application priority of the application to be recommended, which is more than or equal to, to be preset using priority threshold value, Xiang Suoshu first is used
Recommend the application to be recommended in family;Or
When the User Priority of the application to be recommended is more than or equal to pre-set user priority threshold value, determines and used to second
Recommend first application in family.
12. terminal described in 0 or 11 according to claim 1, which is characterized in that first attributive character is the first user property
Feature, second attributive character are the second application attribute feature;
The characteristic processing unit is used for:
To include in the first user property feature of feature acquiring unit acquisition and application recommendation network model set
It is each matched using the user property feature of recommendation network model interaction, from the application recommendation network model set really
It makes and applies recommendation network model associated by corresponding first user of the first user property feature;
Wherein, described using the user property in recommendation network model set further including other users except first user
Other application recommendation network model associated by feature, the other application recommendation network model are associated by the other users
The training of sample application feature obtains.
13. terminal described in 0 or 11 according to claim 1, which is characterized in that first attributive character is the first application attribute
Feature, second attributive character are second user attributive character;
The characteristic processing unit is used for:
To include in the first application attribute feature of feature acquiring unit acquisition and application recommendation network model set
It is each matched using the application attribute feature of recommendation network model interaction, from the application recommendation network model set really
It makes the first application attribute feature corresponding described first and applies recommendation network model using associated;
Wherein, it is described using further include in recommendation network model set it is described first application except other application application attribute
Other application recommendation network model associated by feature, the other application recommendation network model are associated by the other application
The training of sample of users feature obtains.
14. a kind of computer storage medium, which is characterized in that the computer storage medium is stored with a plurality of instruction, when described
When instruction is run at the terminal, so that the terminal executes method as claimed in any one of claims 1 to 9 wherein.
15. a kind of terminal characterized by comprising processor and memory;Wherein, the memory is stored with computer journey
Sequence, the computer program are suitable for being loaded by the processor and executing method as claimed in any one of claims 1 to 9 wherein.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810301039.5A CN110351318B (en) | 2018-04-04 | 2018-04-04 | Application recommendation method, terminal and computer storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810301039.5A CN110351318B (en) | 2018-04-04 | 2018-04-04 | Application recommendation method, terminal and computer storage medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110351318A true CN110351318A (en) | 2019-10-18 |
CN110351318B CN110351318B (en) | 2022-04-29 |
Family
ID=68172811
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810301039.5A Active CN110351318B (en) | 2018-04-04 | 2018-04-04 | Application recommendation method, terminal and computer storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110351318B (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2020108605A1 (en) * | 2018-11-29 | 2020-06-04 | 腾讯科技(深圳)有限公司 | Recommendation method, apparatus, and storage medium |
CN111277706A (en) * | 2020-01-06 | 2020-06-12 | 宇龙计算机通信科技(深圳)有限公司 | Application recommendation method and device, storage medium and electronic equipment |
CN112269928A (en) * | 2020-10-23 | 2021-01-26 | 百度在线网络技术(北京)有限公司 | User recommendation method and device, electronic equipment and computer readable medium |
CN112559856A (en) * | 2020-12-04 | 2021-03-26 | 中国联合网络通信集团有限公司 | Application sorting method and device |
CN112883275A (en) * | 2021-03-17 | 2021-06-01 | 北京乐我无限科技有限责任公司 | Live broadcast room recommendation method, device, server and medium |
CN113297417A (en) * | 2020-02-24 | 2021-08-24 | 北京达佳互联信息技术有限公司 | Video pushing method and device, electronic equipment and storage medium |
CN113505157A (en) * | 2021-07-08 | 2021-10-15 | 深圳市研强物联技术有限公司 | IoT cloud-based wearable device pairing method and system |
CN113573106A (en) * | 2020-04-28 | 2021-10-29 | 北京达佳互联信息技术有限公司 | Model updating method and device for multimedia information and server |
WO2022016556A1 (en) * | 2020-07-24 | 2022-01-27 | 华为技术有限公司 | Neural network distillation method and apparatus |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103890812A (en) * | 2011-08-30 | 2014-06-25 | 诺基亚公司 | Method and apparatus for providing recommendations based on context-aware group modeling |
CN105183882A (en) * | 2015-09-23 | 2015-12-23 | 百度在线网络技术(北京)有限公司 | Application software recommending method and device |
CN105933425A (en) * | 2016-05-18 | 2016-09-07 | 北京奇虎科技有限公司 | Application recommendation method and device |
CN105956146A (en) * | 2016-05-12 | 2016-09-21 | 腾讯科技(深圳)有限公司 | Article information recommending method and device |
CN107578332A (en) * | 2017-09-22 | 2018-01-12 | 深圳乐信软件技术有限公司 | A kind of method, apparatus, equipment and storage medium for recommending cash commodity |
CN107766561A (en) * | 2017-11-06 | 2018-03-06 | 广东欧珀移动通信有限公司 | Method, apparatus, storage medium and the terminal device that music is recommended |
-
2018
- 2018-04-04 CN CN201810301039.5A patent/CN110351318B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103890812A (en) * | 2011-08-30 | 2014-06-25 | 诺基亚公司 | Method and apparatus for providing recommendations based on context-aware group modeling |
CN105183882A (en) * | 2015-09-23 | 2015-12-23 | 百度在线网络技术(北京)有限公司 | Application software recommending method and device |
CN105956146A (en) * | 2016-05-12 | 2016-09-21 | 腾讯科技(深圳)有限公司 | Article information recommending method and device |
CN105933425A (en) * | 2016-05-18 | 2016-09-07 | 北京奇虎科技有限公司 | Application recommendation method and device |
CN107578332A (en) * | 2017-09-22 | 2018-01-12 | 深圳乐信软件技术有限公司 | A kind of method, apparatus, equipment and storage medium for recommending cash commodity |
CN107766561A (en) * | 2017-11-06 | 2018-03-06 | 广东欧珀移动通信有限公司 | Method, apparatus, storage medium and the terminal device that music is recommended |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2020108605A1 (en) * | 2018-11-29 | 2020-06-04 | 腾讯科技(深圳)有限公司 | Recommendation method, apparatus, and storage medium |
US11709902B2 (en) | 2018-11-29 | 2023-07-25 | Tencent Technology (Shenzhen) Company Limited | Recommendation method and apparatus, and storage medium |
CN111277706A (en) * | 2020-01-06 | 2020-06-12 | 宇龙计算机通信科技(深圳)有限公司 | Application recommendation method and device, storage medium and electronic equipment |
CN111277706B (en) * | 2020-01-06 | 2022-09-02 | 宇龙计算机通信科技(深圳)有限公司 | Application recommendation method and device, storage medium and electronic equipment |
CN113297417B (en) * | 2020-02-24 | 2024-05-07 | 北京达佳互联信息技术有限公司 | Video pushing method, device, electronic equipment and storage medium |
CN113297417A (en) * | 2020-02-24 | 2021-08-24 | 北京达佳互联信息技术有限公司 | Video pushing method and device, electronic equipment and storage medium |
CN113573106B (en) * | 2020-04-28 | 2023-03-21 | 北京达佳互联信息技术有限公司 | Model updating method and device for multimedia information and server |
CN113573106A (en) * | 2020-04-28 | 2021-10-29 | 北京达佳互联信息技术有限公司 | Model updating method and device for multimedia information and server |
WO2022016556A1 (en) * | 2020-07-24 | 2022-01-27 | 华为技术有限公司 | Neural network distillation method and apparatus |
CN112269928A (en) * | 2020-10-23 | 2021-01-26 | 百度在线网络技术(北京)有限公司 | User recommendation method and device, electronic equipment and computer readable medium |
CN112559856B (en) * | 2020-12-04 | 2023-06-02 | 中国联合网络通信集团有限公司 | Application ordering method and device |
CN112559856A (en) * | 2020-12-04 | 2021-03-26 | 中国联合网络通信集团有限公司 | Application sorting method and device |
CN112883275A (en) * | 2021-03-17 | 2021-06-01 | 北京乐我无限科技有限责任公司 | Live broadcast room recommendation method, device, server and medium |
CN112883275B (en) * | 2021-03-17 | 2024-01-19 | 北京乐我无限科技有限责任公司 | Live broadcast room recommendation method, device, server and medium |
CN113505157A (en) * | 2021-07-08 | 2021-10-15 | 深圳市研强物联技术有限公司 | IoT cloud-based wearable device pairing method and system |
CN113505157B (en) * | 2021-07-08 | 2023-10-20 | 深圳市研强物联技术有限公司 | Wearable device pairing method and system based on internet of things (IoT) cloud |
Also Published As
Publication number | Publication date |
---|---|
CN110351318B (en) | 2022-04-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110351318A (en) | Using the method, terminal and computer storage medium of recommendation | |
US10958748B2 (en) | Resource push method and apparatus | |
CN111680221B (en) | Information recommendation method, device, equipment and computer readable storage medium | |
CN111061946B (en) | Method, device, electronic equipment and storage medium for recommending scenerized content | |
CN110046304A (en) | A kind of user's recommended method and device | |
CN110008375A (en) | Video is recommended to recall method and apparatus | |
CN106575503A (en) | Session context modeling for conversational understanding systems | |
CN110245301A (en) | A kind of recommended method, device and storage medium | |
Huang et al. | An efficient hybrid recommendation model with deep neural networks | |
CN110364146A (en) | Audio recognition method, device, speech recognition apparatus and storage medium | |
CN112380453B (en) | Article recommendation method and device, storage medium and equipment | |
CN110413888B (en) | Book recommendation method and device | |
CN110147882A (en) | Training method, crowd's method of diffusion, device and the equipment of neural network model | |
CN110020175B (en) | Search processing method, processing equipment and system | |
CN111241394A (en) | Data processing method and device, computer readable storage medium and electronic equipment | |
Shiue et al. | Development of machine learning‐based real time scheduling systems: using ensemble based on wrapper feature selection approach | |
CN110347781A (en) | Article falls discharge method, article recommended method, device, equipment and storage medium | |
CN110162609A (en) | For recommending the method and device asked questions to user | |
CN117009650A (en) | Recommendation method and device | |
CN110110218A (en) | A kind of Identity Association method and terminal | |
CN113656699B (en) | User feature vector determining method, related equipment and medium | |
CN116823410B (en) | Data processing method, object processing method, recommending method and computing device | |
CN109558528A (en) | Article method for pushing, device, computer readable storage medium and server | |
CN115329183A (en) | Data processing method, device, storage medium and equipment | |
CN112052386A (en) | Information recommendation method and device and storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |