CN105989074B - A kind of method and apparatus recommend by mobile device information cold start-up - Google Patents
A kind of method and apparatus recommend by mobile device information cold start-up Download PDFInfo
- Publication number
- CN105989074B CN105989074B CN201510070689.XA CN201510070689A CN105989074B CN 105989074 B CN105989074 B CN 105989074B CN 201510070689 A CN201510070689 A CN 201510070689A CN 105989074 B CN105989074 B CN 105989074B
- Authority
- CN
- China
- Prior art keywords
- user
- app
- mobile device
- recommendation
- interest tags
- 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.)
- Active
Links
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Information Transfer Between Computers (AREA)
Abstract
A kind of method and apparatus recommend by mobile device information cold start-up, this method comprises the following steps: obtaining the mobile device information of user, obtains all APP information of the mobile device model and installation of the user on the mobile device by the operating system of the mobile device of the user;Generate the recommendation list for being directed to the user, comprising: the first recommendation list is generated based on collaborative filtering, the content that other users similar with the APP of the mobile device model of the user and/or installation in database are liked generates recommendation list as recommendation;Or recommendation list is generated based on interest tags mapping, App is explicitly mapped to one or more interest tags, then screens corresponding content as recommendation according to each interest tags and generate the second recommendation list;Give the commending contents in first recommendation list or second recommendation list to the user.The invention also discloses the devices for recommend by mobile device information cold start-up.
Description
Technical field
The present invention relates to a kind of method and apparatus for recommend by mobile device information cold start-up.
Background technique
The cold start-up problem of recommender system refers to and lacks enough data for new custom system to capture the interest of user
And effectively recommendation.This problem is a significant challenge of the recommender system in actual product application.It is directed to numerous
In this solution to the problem, a kind of widely used method is to encourage user with social networks (Social Network
Service:SNS) account logs in recommender system.Recommender system can use user social contact network information (such as concern relation,
Friend relation, interest tags, publication content etc.) initialising subscriber interest model, to effectively be recommended.Application No. is
" 201410020292 ", the China of entitled " microblogging word cloud generation method and access support system based on Users' Interests Mining "
Application for a patent for invention proposes a kind of method of the keyword as user interest keyword by excavating user's publication content.
For another example it is published in " application of the label propagation algorithm in microblog users interest graph " text of in July, 2012 " programmer " magazine,
A kind of social networks using user are described to converge to user the interest tags of user good friend or perpetual object
Method facilitates the sparsity for solving the problems, such as user's own interests label in this way, because of all not interested label of a large number of users.
But in practical applications, many users are because worry privacy concern or dislike trouble to select to use social networks account
Number log in recommended products.According to statistics, recommend in class product in the news of some hot topics, with microblogging, QQ, wechat etc. in all users
What mainstream social networks account logged in is no more than 50%.This allows for existing method and is difficult thoroughly to solve the new user's of recommender system
Cold start-up problem.In contrast, the invention proposes a kind of method universality is stronger, and the cold starting effect on many users
And no less than the excavation user social contact network information.
Summary of the invention
The technical problem to be solved by the present invention is in view of the above drawbacks of the prior art, provide one kind to pass through mobile device
The method and apparatus that information recommend cold start-up, excavate user interest using user's mobile device information, to thoroughly solve
The certainly cold start-up problem of the new user of recommender system.
To achieve the goals above, the present invention provides a kind of sides for recommend by mobile device information cold start-up
Method, wherein include the following steps:
S100, the mobile device information for obtaining user, obtain the user by the operating system of the mobile device of the user
Mobile device model and installation all APP information on the mobile device;
S200, the recommendation list for being directed to the user is generated, comprising:
S201, based on collaborative filtering generate the first recommendation list, by database with the mobile device model of the user and/
Or the content that the similar other users of APP of installation are liked is as recommendation the first recommendation list of generation;Or
S202, the second recommendation list is generated based on interest tags mapping, App is explicitly mapped to one or more interest
Then label screens corresponding content as recommendation according to each interest tags and generates the second recommendation list;
S300, the commending contents in first recommendation list or the second recommendation list are given to the user.
The above-mentioned method for recommend by mobile device information cold start-up, wherein based on association in the step S201
It is same to be filtered into the collaborative filtering method based on similar users discovery, comprising:
S2011, screening is common distinction App;
S2012, common smart phone type is chosen;
S2013, App and type are mapped to specific dimension;
S2014, for giving user, extraction mobile device feature vector is given in the mobile device information of user from this;
S2015, vector distance and the nearest K user of the given user are found based on WeakAND algorithm;
S2016, the highest content of clicking rate is counted in this K user as recommendation.
The above-mentioned method for recommend by mobile device information cold start-up, wherein extracted in the step S2014
The mobile device feature vector of given user includes:
S20141, each App is mapped to the dimension between one [0, N-1];
S20142, each type is mapped to the dimension between one [N, M-1], the value of dimension corresponding with user's type is
1, the value of other dimensions is 0;
S20143, user installation App correspond to dimension value be the nearest specified number of days of the user access times, other dimension
The value of degree is 0.
The above-mentioned method for recommend by mobile device information cold start-up, wherein based on association in the step S201
It is same to be filtered into the collaborative filtering method based on APP and model information, comprising:
S201a, the high click that clicking rate in the user group of each App and common type is Top K is periodically counted respectively
Contents list;
S201b, given user's mobile device, the App and type installed by user's mobile device obtain correspondence respectively
Described high click contents list;
S201c, the height for merging acquisition according to the weight of corresponding App and type click contents list and will be therein
Top N is as recommendation.
The above-mentioned method for recommend by mobile device information cold start-up, wherein in the step S201c, merge
The sequence of recommendation i is calculated using following formula when the high click contents list:
Score (i)=sum (wa)+wd
Wherein, a represents the App of user, and i is appeared in the Top K high click contents list of a, and wa represents App
A can represent the weight of user interest;D represents the type of user, and wd, which represents user's type, can represent the weight of user interest.
The above-mentioned method for recommend by mobile device information cold start-up, wherein the step S202 is based on interest
Label mapping generates recommendation list
S2021, interest tags are marked to App;
S2022, given user's mobile device, obtain the interest tags of the corresponding A pp of user's mobile device installation;
S2023, will content relevant to the interest tags as recommendation.
The above-mentioned method for recommend by mobile device information cold start-up, wherein marked in the step S2021
Interest tags include:
S20211, the tag database for establishing recommender system itself;
S20212, the label that each APP in APP application market is grabbed with webpage capture technology;
S20213, the label mapping of the APP of crawl into the tag database.
The above-mentioned method for recommend by mobile device information cold start-up, wherein marked in the step S2021
Interest tags include:
S2021a, the global interest tags Top M for counting all users;
S2021b, an App is given, statistics is mounted in the user group of the App, most popular Top L interest mark
Label;
S2021c, compare the Top L interest tags and global interest tags Top M, taking-up is different from overall situation Top M
Interest tags of the interest tags as the App.
The above-mentioned method for recommend by mobile device information cold start-up, wherein counted in the step S2021a
The global interest tags Top M of all users calculates the interest tags vector of user using following formula:
Wherein, Ti represents the interest tags vector of i-th of user action, and wi represents the weight of i-th of user action.
In order to which above-mentioned purpose is better achieved, the present invention also provides one kind to carry out recommending cold open by mobile device information
Dynamic method, wherein include the following steps:
S100, the mobile device information for obtaining user, obtain the user by the operating system of the mobile device of the user
Mobile device model and installation all APP information on the mobile device;
S200, the first recommendation list to the user is generated based on collaborative filtering, by the movement in database with the user
The content that device model and/or the similar other users of the APP of installation are liked generates the first recommendation list as recommendation;
S300, the second recommendation list to the user is generated based on interest tags mapping, App is explicitly mapped to one
Or multiple interest tags, corresponding content then, which is screened, as recommendation according to each interest tags generates the second recommendation column
Table;
S400, merge first recommendation list and second recommendation list, and according to wherein listed recommendation
Weighted sum rearranges recommendation order, generates a preferred recommendation list;
S500, the commending contents in the preferred recommendation list are given to the user.
In order to which above-mentioned purpose is better achieved, the present invention also provides one kind for carrying out above by mobile device information
Recommend the device of the method for cold start-up.
Beneficial functional of the invention is:
It carries out that cold start-up is recommended to compare using the user social contact network information in the prior art, the present invention directly utilizes user
The information of mobile device is not necessarily to user's Unsolicited Grant, coverage rate 100%.And in practical applications, effect of the invention will be excellent
In the recommendation cold start-up based on the user social contact network information, the either use time of user and content clicking rate all reaches or surpasses
Same level is crossed.
Below in conjunction with the drawings and specific embodiments, the present invention will be described in detail, but not as a limitation of the invention.
Detailed description of the invention
Fig. 1 is the method flow diagram of one embodiment of the invention;
Fig. 2 is the collaborative filtering method flow chart based on similar users discovery of one embodiment of the invention;
Fig. 3 is the extraction mobile device feature vector schematic diagram of one embodiment of the invention;
Fig. 4 is the collaborative filtering method flow chart based on APP and model information of one embodiment of the invention;
Fig. 5 is that being mapped based on interest tags for one embodiment of the invention generates recommendation list flow chart;
Fig. 6 is the method flow diagram of another embodiment of the present invention.
Wherein, appended drawing reference
S100-S500, S2011-S2016, S201a-S201c, S2021-S2023 step
Specific embodiment
Structural principle and working principle of the invention are described in detail with reference to the accompanying drawing:
It is the method flow diagram of one embodiment of the invention referring to Fig. 1, Fig. 1.Of the invention is carried out by mobile device information
The method for recommending cold start-up can be used for recommending any pair that can be classified with interest tags such as article, picture, video, commodity
As including the following steps:
Step S100, the mobile device information for obtaining user, being obtained by the operating system of the mobile device of the user should
All APP information of the mobile device model and installation of user on the mobile device, such as mobile device operation system can be passed through
The API that system provides obtains device model, the App installed, the mobile devices information such as App being currently running.It leads currently on the market
The Mobile operating system of stream has disclosed API (Application Program Interface: application programming interfaces) confession
Third party App obtains some basic facility informations.After obtaining user's authorization, third party App obtains the access right of these API
Limit obtains the information needed.For example, in Android platform, all APP information of the available device model of third party App and installation
(installation kit name).On iOS platform, situation is somewhat more complex, and third party App cannot be directly acquired to be installed in user's mobile device
App list, but this information can be obtained indirectly by way of inquiring the equipment and whether some App is installed.This method
It is required that commonly using the id list of iOS App built in third party App, these data are can to grab from the webpage version of App Store
's.
Step S200, the recommendation list for being directed to the user is generated, is further included steps of
Step S201, based on collaborative filtering generate recommendation list, by database with the mobile device model of the user and/
Or the content that the similar other users of APP of installation are liked generates recommendation list as recommendation, i.e., by device model, pacifies
The same device models of information recommendations such as App are filled, the content that same App user clicks, shares, collects is installed.This strategy should not
It asks recommender system to understand user's type and the user interest that App reflects is installed, the user U given for one, it is only necessary to push away
Recommend the content that user identical with U type or that installation App is similar likes.Specifically, there are two types of common for this strategy
Implementation method.
Alternatively, step S202 can also be used, recommendation list and Synergistic method difference are generated based on interest tags mapping, this
Kind method can only utilize App information, cannot utilize model information.App is explicitly mapped to one or more interest tags, so
Corresponding content is screened as recommendation according to each interest tags afterwards and generates recommendation list, i.e., by establishing App and interest
The mapping relations of label, the user for installing specific App utilize the interest tags recommendation of mapping.For example, for App
" taking journey ", if this App is stamped the labels such as " travel for commercial purpose ", " workplace ", (corresponding user is likely to one and often goes out
The business people of difference), for being mounted with the user of this App, we can recommend that " travel for commercial purpose " and " workplace " is marked
The popular article of label.Automatically there is the application of comparative maturity to the method that article labels, so it will not be repeated.
Step S300, the commending contents in the recommendation list are given to the user.
Referring to fig. 2, Fig. 2 is the collaborative filtering method flow chart based on similar users discovery of one embodiment of the invention.Institute
Stating in step S201 can be for example the collaborative filtering method found based on similar users, give a user U, close to obtain
Suitable content recommendation lists IU, needs first to find type and App installs the list user similar with U.Specifically can include:
Step S2011, screening is common distinction App, such as can be by collecting the sufficiently large application of some user volume
The APP mount message of all users, counts the user installation ratio of each APP, which can simulate all App in institute substantially
There is the proportion of installation in mobile interchange network users.Firstly, the filtering proportion of installation is greater than 50% App, it is this kind of substantially micro-
Rich, wechat, this whole people App of Baidu's client, to identification user personalized interest without too big help.Secondly, filtering installation
App of the ratio less than 1%, it is assumed that mobile interchange network users have 500,000,000, and 1% means 50,000 installation, can lower than this number
To think excessively minority, without statistical significance;
Step S2012, common smart phone type is chosen;
Step S2013, App and type are mapped to specific dimension;
Step S2014, for give user, given from this in mobile device information of user extract mobile device feature to
Amount;
Step S2015, vector distance is found based on WeakAND algorithm and this gives K nearest user of user;
Step S2016, the highest content of clicking rate is counted in this K user as recommendation.
Wherein, the mobile device feature vector that given user is extracted in the step S2014, can be the shifting of each user
Dynamic facility information is expressed as the equipment feature vector VU of N+M dimension, and wherein N represents considered App type, and M represents mainstream
The quantity of smart phone type.It specifically includes:
Step S20141, the dimension (such as ctrip.com-> 2340) being mapped to each App between one [0, N-1];
Step S20142, each type is mapped to the dimension (such as millet 2-> 10012) between one [N, M-1],
The information of a mobile device given in this way, so that it may a corresponding vector is obtained, dimension corresponding with user's type
Value is 1, and the value of other dimensions is 0;
Step S20143, it is the nearest specified number of days of the user (such as nearest seven that the App of user installation, which corresponds to the value of dimension,
It) access times, the value of other dimensions is 0.
For example, for being mounted with " ctrip.com (2340) ", millet 2 (10012) equipment of " understanding ball Supreme Being (3078) ",
Vector described in Fig. 3 can be obtained.Wherein, it is 1 that the value of 0-2339 dimension, which is the value of the 0, the 2340th dimension, 2341-
The value of 3077 dimensions is that the value of the 0, the 3078th dimension is 3, and the value of 3079-10011 dimension is that the value of the 0, the 10012nd dimension is 1,
The value of 10013- (N+M-1) dimension is 0, then the value of the 2340th dimension be the value of the 1, the 3078th dimension be 3 respectively represent it is corresponding
Access times in APP seven days, the value of the 10012nd dimension are 1 to represent user's type as 10012 (corresponding millets 2).I.e. at this
On a vector, the value that user installation App corresponds to dimension is indicated with the access times of user, and the value that user's type corresponds to dimension is 1,
The value of other dimensions is all 0.
The equipment feature vector VU of a given user find with can be convenient the Top K most like with this vector to
Amount.If doing a calculating to all devices feature vector, calculation amount is very big, and (for example the recommended products user of some mainstreams is
By hundred million), can not only consume a large amount of computing resources, and postpone possibly not receiving and (need several hours).This problem warp
The WeakAND algorithm of allusion quotation can well solve, and minute grade can be disposed in practice.After finding similar user, Ke Yi
In these users in statistics several hours in the past the highest content of clicking rate as recommendation list IU.It should be noted that currently,
The mobile application market of mainstream often has hundreds of thousands kind using available for download.That is, theoretically the N in this method can be with
Reach the magnitude of hundreds of thousands, the expense of calculating similar users in this way can be bigger.In practical applications, in order to remove noise and drop
Low computing cost can filter out the application of those downloads king-sized whole people grade (such as wechat, microblogging, QQ, Baidu map
Deng) and the especially few application of download, the App type for actually participating in calculating is for example 10,000 or so.
Assuming that the equipment feature vector of user u is Vu, Top K vector is exactly the feature vector most like with Vu.And vector
Similitude can generally be measured using the angle of vector distance or vector.WeakAND algorithm is conventionally used to examine in text
Quickly search the most like Top K term vector (similar document) of term vector (inquiry document) in rope field.Its main thought
It is that the vector distance of given a term vector V1, another term vector V2 and V1 have a upper limit, this upper limit can decompose
For the sum of the contribution upper limit of each word.The contribution upper limit of each word can precalculate, for example a word " national football team " is in institute
Maximum weight is 0.2 inside article term vector, then the contribution upper limit of this word is exactly 0.2*0.2=0.04, that is,
It says, is easy to be calculated, two articles comprising a common word " national football team ", vector distance is up to 0.04.It is fast according to this
The method that speed calculates the vector distance upper limit gives V1, to find the nearest vector of Top K vector distance, does not just have to all
Candidate vector all calculates a distance and then sorts.It can usually safeguard that the queue that a length is K can for preceding K vector
To be directly first put into queue.Since the K+1 vector, neutralizing apart from whether the upper limit is less than queue for each vector is checked
V1, if it is lower, just directly abandoning, otherwise calculates actual distance apart from the smallest vector again, sees whether that being less than queue neutralizes
V1 is apart from the smallest vector.If it is less than or be equal to, just abandon, if it does, just in queue and V1 apart from the smallest vector
Replacement comes out.In this way for many vectors, so that it may save the calculating of distance, and only need to scan candidate vector one time.
In the present embodiment, if App and mobile device model are treated as word, user's mobile device feature vector treats as term vector, just
Can borrow WeakAND algorithm rapidly obtain and the other users of the most like K of user's mobile device vector movement set
Standby feature vector, i.e. Top K vector.
Referring to fig. 4, Fig. 4 is the collaborative filtering method flow chart based on APP and model information of one embodiment of the invention.Institute
State the collaborative filtering method that also may be based on APP and model information in step S201 based on collaborative filtering, comprising:
Step S201a, periodically (such as each hour) counts the corresponding user of each App respectively and each commonly uses type
Clicking rate is that the high of Top K clicks contents list within the past period in user group, and then statistical result is stored in
In database or cache server memory-based (access speed is faster);
Step S201b, user's mobile device is given, the App and type installed by user's mobile device is obtained respectively
The corresponding high click contents list;
Step S201c, merge according to the weight of corresponding App and type obtain described it is high click contents list and by its
In Top N as recommendation.
Wherein, in the step S201c, it may be considered that the weight of different App pushes away when merging the high click contents list
The sequence for recommending content i is calculated using following formula:
Score (i)=sum (wa)+wd
Wherein, a represents the App of user, and i is appeared in the Top K high click contents list of a, and wa is represented should
App a can represent the weight of user interest;D represents the type of user, and wd, which represents user's type, can represent user interest
Weight.With the score according to each content, K contents of Top can be chosen and returned as recommendation results.
It is that being mapped based on interest tags for one embodiment of the invention generates recommendation list flow chart referring to Fig. 5, Fig. 5.It is described
Step S202, which is based on interest tags mapping generation recommendation list, may include steps of:
Step S2021, interest tags are marked to App;
Step S2022, user's mobile device is given, the interest of the corresponding A pp of user's mobile device installation is obtained
Label;
Step S2023, will content relevant to the interest tags as recommendation.
Wherein, mark interest tags that following method can be used in the step S2021, comprising:
Step S20211, the tag database of recommender system itself is established;
Step S20212, with the label of each APP in webpage capture technology crawl APP application market (such as pea pods);
Step S20213, the label mapping of the APP of crawl into the tag database.Why need to map
It is because the label system of App application market may be not consistent with recommender system itself.Such as a UEFA Champions League video App,
The label of possible only one " UEFA Champions League " in application market, and may there is no " UEFA Champions League " inside recommender system but have " international
The label of football ", " UEFA Champions League " should just be mapped as " international soccer " in this case.Mapping relations can by manual maintenance,
It can be safeguarded by system, since it is desired that the number of labels of mapping is usually no more than 1,000, and not needed to each App automatically
There are more understandings, is more easier so operating.
It is a kind of calculate App weight method be to compare the Top K list and whole station clicking rate Top K content of an App
Difference illustrates that this App gets over the personalized interest that can embody user, weight is bigger if two list differences are bigger.It is similar
, the weight for obtaining type can also be compared according to the Top K list and whole station clicking rate Top K content of type.For example,
In the user of today's tops, it is mounted with that the user of " knowing daily paper " likes the article clicked and the highest article of whole station clicking rate and has
Significant difference.The highest article of whole station clicking rate is often the article for entertaining Eight Diagrams and social intriguing story class, and is mounted with this App
User then preference minority high-quality content.Therefore, the interest expression weight of this App is compared with general instrumental App
It is very high.Specifically, wa can be calculated with following formula
Wa=1- | Ia I |/K
Wherein Ia, I respectively represent the Top K clicking rate contents list for being mounted with App a user and whole station clicking rate Top K
Contents list, | Ia I | indicate the size of both of which intersection.It similarly, can also be according to the Top K list of type and whole station
Clicking rate Top K content compares the weight wd for obtaining type.
Therefore, mark interest tags that following method can also be used in the step S2021, comprising:
Step S2021a, the global interest tags Top M of all users is counted;
Step S2021b, an App is given, statistics is mounted in the user group of the App, and most popular Top L emerging
Interesting label;
Step S2021c, compare the Top L interest tags and global interest tags Top M, taking-up is different from the overall situation
Interest tags of the interest tags of Top M as the App.
This method requires the interest tags for excavating each user in advance, then gives an App, and statistics is mounted with this
In the user group of App, most popular Top K interest tags.Finally take this Top K interest tags and all users
The Top M interest tags (global Top M) counted compare, and take out and overall situation Top M is different as the emerging of this App
Interesting label.In the case where each single item recommendation has interest tags, excavating user interest label by user behavior is also
One more mature technology.It is general to use the behavioral data of recommendation service (for example, browsing in which by user in recording station
Hold, which content click/collection/commented on), and the interest tags of user and every are excavated according to the interest tags of content
The weight of a label, the two together constitute the interest tags vector of user.The specific method is as follows:
A. a weight w is set for every kind of user action act, for example clicks 1 point of calculation, browsing is still without clicking calculation-
0.2 point, collection calculates 5 points;
B. sequence of user actions [act1, act2 ..., act3] is given, the interest tags vector of user calculates as follows:
Wherein Ti represents the interest tags vector of i-th of user action, and wi represents the weight of i-th of user action.
It is the method flow diagram of another embodiment of the present invention referring to Fig. 6, Fig. 6.In the present embodiment, believed by mobile device
The method that breath recommend cold start-up, it may include following steps:
Step S100, the mobile device information for obtaining user, being obtained by the operating system of the mobile device of the user should
All APP information of the mobile device model and installation of user on the mobile device;
Step S200, generated based on collaborative filtering to the first recommendation list of the user, by database with the user's
The content that mobile device model and/or the similar other users of the APP of installation are liked generates first as recommendation and recommends column
Table;
Step S300, the second recommendation list to the user is generated based on interest tags mapping, App is explicitly mapped to
Then one or more interest tags screen corresponding content as recommendation according to each interest tags and generate the second recommendation
List;
Step S400, merge first recommendation list and second recommendation list, and according in wherein listed recommendation
The weighted sum of appearance rearranges recommendation order, generates a preferred recommendation list;
Step S500, the commending contents in the preferred recommendation list are given to the user.
This method fully utilizes two methods of the advantages of based on collaborative filtering and based on interest tags mapping, both mixing
Recommendation results.Specifically, the score of a recommendation can use the weighted sum for being expressed as the score in two kinds of strategies, in this way
The content that the two tends to recommend will obtain higher weight and preferential recommendation to user.It is substantially exactly one two and obtains
The consolidation problem for dividing list can have many methods in processing.For example, enabling the score of a content:
Score (i)=w1*score_1 (i)+w2*score_2 (i)
Wherein score_1 indicates i in the score of list 1, and score_2 indicates i in the score of list 2, and weight w1 and w2 can
To be come out by experimental debugging.
The present invention also provides a kind of for recommend above by mobile device information the device of the method for cold start-up, packet
Include data acquisition module, data processing module and data outputting module, wherein data acquisition module is used to obtain the movement of user
Facility information obtains the mobile device model of the user by the operating system of the mobile device of the user and is mounted on the movement
All APP information in equipment;Data processing module is used to handle the mobile device information of the data collecting module collected simultaneously
Generate recommendation list, including being generated based on collaborative filtering to the first recommendation list of the user, by database with the user's
The content that mobile device model and/or the similar other users of the APP of installation are liked generates first as recommendation and recommends column
Table;Or the second recommendation list to the user is generated based on interest tags mapping, App is explicitly mapped to one or more
Then interest tags screen corresponding content as recommendation according to each interest tags and generate the second recommendation list;And
Merge first recommendation list and second recommendation list, and is rearranged according to the weighted sum of wherein listed recommendation
Recommendation order generates a preferred recommendation list;In the recommendation list that data outputting module is generated for propelling data processing module
Preferred recommendation to relative users.
The present invention directly utilizes the information of user's mobile device, is not necessarily to user's Unsolicited Grant, coverage rate 100%.And
In practical application, effect of the invention is better than the recommendation cold start-up based on the user social contact network information, the either use of user
Time and content clicking rate are all up higher at least same level.
Certainly, the present invention can also have other various embodiments, without deviating from the spirit and substance of the present invention, ripe
It knows those skilled in the art and makes various corresponding changes and modifications, but these corresponding changes and change in accordance with the present invention
Shape all should fall within the scope of protection of the appended claims of the present invention.
Claims (9)
1. a kind of method for carrying out commending contents cold start-up by mobile device information, which comprises the steps of:
S100, the mobile device information for obtaining user, the shifting of the user is obtained by the operating system of the mobile device of the user
All APP information of dynamic device model and installation on the mobile device;
S200, the first recommendation list to the user is generated based on collaborative filtering, by the mobile device in database with the user
The content that model and/or the similar other users of the APP of installation are liked generates the first recommendation list as recommendation;
S300, the second recommendation list to the user is generated based on interest tags mapping, App is explicitly mapped to one or more
Then a interest tags screen corresponding content as recommendation according to each interest tags and generate the second recommendation list;
S400, merge first recommendation list and second recommendation list, and according to the weighting of wherein listed recommendation
With rearrange recommendation order, generate a preferred recommendation list;
S500, the commending contents in the preferred recommendation list are given to the user.
2. the method for carrying out commending contents cold start-up by mobile device information as described in claim 1, which is characterized in that institute
Stating in step S200 based on collaborative filtering is the collaborative filtering method found based on similar users, comprising:
S2001, screening is common distinction App;
S2002, common smart phone type is chosen;
S2003, App and type are mapped to specific dimension;
S2004, for giving user, extraction mobile device feature vector is given in the mobile device information of user from this;
S2005, vector distance and the nearest K user of the given user are found based on WeakAND algorithm;
S2006, the highest content of clicking rate is counted in this K user as recommendation.
3. the method for carrying out commending contents cold start-up by mobile device information as claimed in claim 2, which is characterized in that institute
It states and extracts the mobile device feature vector of given user in step S2004 and include:
S20041, each App is mapped to the dimension between one [0, N-1];
S20042, each type being mapped to the dimension between one [N, M-1], the value of dimension corresponding with user's type is 1,
The value of his dimension is 0;
S20043, user installation App correspond to dimension value be the nearest specified number of days of the user access times, other dimensions
Value is 0.
4. the method for carrying out commending contents cold start-up by mobile device information as described in claim 1, which is characterized in that institute
State in step S200 based on collaborative filtering be the collaborative filtering method based on APP and model information, comprising:
S200a, the high click on content that clicking rate is Top K in the user group of each App and common type is periodically counted respectively
List;
S200b, given user's mobile device, the App and type installed by user's mobile device obtain corresponding institute respectively
State high click contents list;
S200c, merge that obtain described is high to click contents list and by Top therein according to the weight of corresponding App and type
N is as recommendation.
5. the method for carrying out commending contents cold start-up by mobile device information as claimed in claim 4, which is characterized in that institute
It states in step S200c, the sequence of recommendation i is calculated using following formula when merging the high click contents list:
Score (i)=sum (wa)+wd
Wherein, a represents the App of user, and i is appeared in the Top K high click contents list of a, and wa represents App a energy
Enough represent the weight of user interest;D represents the type of user, and wd, which represents user's type, can represent the weight of user interest.
6. the method for carrying out commending contents cold start-up by mobile device information as described in claim 1, which is characterized in that institute
Step S300, which is stated, based on interest tags mapping the second recommendation list of generation includes:
S3001, interest tags are marked to App;
S3002, given user's mobile device, obtain the interest tags of the corresponding A pp of user's mobile device installation;
S3003, will content relevant to the interest tags as recommendation.
7. the method for carrying out commending contents cold start-up by mobile device information as claimed in claim 6, which is characterized in that institute
Stating label interest tags in step S3001 includes:
S30011, the tag database for establishing recommender system itself;
S30012, the label that each APP in APP application market is grabbed with webpage capture technology;
S30013, the label mapping of the APP of crawl into the tag database.
8. the method for carrying out commending contents cold start-up by mobile device information as claimed in claim 6, which is characterized in that institute
Stating label interest tags in step S3001 includes:
S3001a, the global interest tags Top M for counting all users;
S3001b, an App is given, statistics is mounted in the user group of the App, most popular Top L interest tags;
S3001c, compare the Top L interest tags and global interest tags Top M, taking-up is different from the emerging of overall situation Top M
Interest tags of the interesting label as the App.
9. a kind of, for being carried out described in any one of the claims 1-8 by mobile device information, commending contents are cold to be opened
The device of dynamic method.
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN2015100674013 | 2015-02-09 | ||
CN201510067401 | 2015-02-09 |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105989074A CN105989074A (en) | 2016-10-05 |
CN105989074B true CN105989074B (en) | 2019-07-05 |
Family
ID=57041126
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510070689.XA Active CN105989074B (en) | 2015-02-09 | 2015-02-11 | A kind of method and apparatus recommend by mobile device information cold start-up |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105989074B (en) |
Families Citing this family (21)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107967276A (en) * | 2016-10-19 | 2018-04-27 | 阿里巴巴集团控股有限公司 | A kind of method and apparatus of recommended |
CN106776859A (en) * | 2016-11-28 | 2017-05-31 | 南京华苏科技有限公司 | Mobile solution App commending systems based on user preference |
CN106846094A (en) * | 2016-12-29 | 2017-06-13 | 广州优视网络科技有限公司 | A kind of method and apparatus for recommending application message based on application has been installed |
CN107223341A (en) * | 2017-04-01 | 2017-09-29 | 深圳市智晟达科技有限公司 | The method and commending system for recommending video are used according to app |
CN109145280B (en) * | 2017-06-15 | 2023-05-12 | 北京京东尚科信息技术有限公司 | Information pushing method and device |
CN108009247B (en) * | 2017-11-30 | 2021-01-12 | 广州酷狗计算机科技有限公司 | Information pushing method and device |
CN108173936A (en) * | 2017-12-27 | 2018-06-15 | 百度在线网络技术(北京)有限公司 | News recommends method and apparatus |
CN108520017B (en) * | 2018-03-21 | 2019-09-10 | Oppo广东移动通信有限公司 | Application program recommended method, device, server and storage medium |
CN108809987B (en) * | 2018-06-14 | 2020-10-23 | 江苏果米文化发展有限公司 | Online game popularization method based on big data analysis |
CN110858231A (en) * | 2018-08-07 | 2020-03-03 | 北京京东尚科信息技术有限公司 | Article recommendation method and device |
CN109410079B (en) * | 2018-09-13 | 2021-08-20 | 清华大学 | Social recommendation method and device under optimal limited attention |
CN110955820A (en) * | 2018-09-22 | 2020-04-03 | 北京微播视界科技有限公司 | Media information interest point recommendation method, device, server and storage medium |
CN111310016B (en) * | 2018-12-11 | 2023-08-04 | 百度在线网络技术(北京)有限公司 | Label mining method, device, server and storage medium |
CN109688458A (en) * | 2019-01-14 | 2019-04-26 | 四川长虹电器股份有限公司 | The implementation method of smart television cloud desktop operation system based on big data algorithm |
CN111814032B (en) * | 2019-04-11 | 2024-05-28 | 阿里巴巴集团控股有限公司 | Cold start recommendation method and device and electronic equipment |
CN110263053B (en) * | 2019-06-17 | 2021-04-27 | 每日互动股份有限公司 | Server for acquiring portrait label of mobile terminal based on mobile terminal data |
CN111708952B (en) * | 2020-06-18 | 2023-10-20 | 小红书科技有限公司 | Label recommending method and system |
CN112596712A (en) * | 2020-12-28 | 2021-04-02 | 上海风秩科技有限公司 | Cold start interface design method, system, electronic equipment and storage medium |
CN113407845A (en) * | 2021-07-14 | 2021-09-17 | 上海明略人工智能(集团)有限公司 | Method and device for information recommendation, electronic equipment and storage medium |
CN113761361A (en) * | 2021-07-29 | 2021-12-07 | 深圳市思为软件技术有限公司 | House property information recommendation method and terminal equipment |
CN114707075B (en) * | 2022-06-06 | 2022-10-25 | 荣耀终端有限公司 | Cold start recommendation method and device |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101959179A (en) * | 2009-07-17 | 2011-01-26 | 华为技术有限公司 | Method for providing mobile terminal application program, and server and mobile terminal |
CN103399967A (en) * | 2013-08-26 | 2013-11-20 | 百度在线网络技术(北京)有限公司 | Software recommending method and system and server |
CN104133878A (en) * | 2014-07-25 | 2014-11-05 | 百度在线网络技术(北京)有限公司 | User label generation method and device |
Family Cites Families (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101937547A (en) * | 2010-09-15 | 2011-01-05 | 宇龙计算机通信科技(深圳)有限公司 | Software and/or software information pushing method, system, acquisition device, software shop service system and mobile terminal |
CN102592223A (en) * | 2011-01-18 | 2012-07-18 | 卓望数码技术(深圳)有限公司 | Commodity recommending method and commodity recommending system |
KR101311511B1 (en) * | 2011-06-17 | 2013-09-25 | 주식회사 케이티 | In-app recommending system and user terminal unit |
CN103354561A (en) * | 2013-06-28 | 2013-10-16 | 贵阳朗玛信息技术股份有限公司 | Pushing method, processing method, and pushing device for configuration information |
CN104216998B (en) * | 2014-09-10 | 2018-03-30 | 广州猎豹网络科技有限公司 | A kind of method, apparatus and terminal device for recommending application program |
CN104239571B (en) * | 2014-09-30 | 2018-04-24 | 北京奇虎科技有限公司 | It is a kind of to carry out using the method and apparatus recommended |
-
2015
- 2015-02-11 CN CN201510070689.XA patent/CN105989074B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101959179A (en) * | 2009-07-17 | 2011-01-26 | 华为技术有限公司 | Method for providing mobile terminal application program, and server and mobile terminal |
CN103399967A (en) * | 2013-08-26 | 2013-11-20 | 百度在线网络技术(北京)有限公司 | Software recommending method and system and server |
CN104133878A (en) * | 2014-07-25 | 2014-11-05 | 百度在线网络技术(北京)有限公司 | User label generation method and device |
Non-Patent Citations (1)
Title |
---|
基于个性化推荐的移动阅读服务***的研究与设计;周伟华;《中国优秀硕士学位论文全文数据库信息科技辑》;20110915(第09期);正文第21-26页 |
Also Published As
Publication number | Publication date |
---|---|
CN105989074A (en) | 2016-10-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105989074B (en) | A kind of method and apparatus recommend by mobile device information cold start-up | |
Johnson | Logistic matrix factorization for implicit feedback data | |
CN103678672B (en) | Method for recommending information | |
CN102591942B (en) | Method and device for automatic application recommendation | |
Kim et al. | Discovering hot topics using Twitter streaming data: social topic detection and geographic clustering | |
CN104077415B (en) | Searching method and device | |
CN106407425A (en) | A method and a device for information push based on artificial intelligence | |
CN108256537A (en) | A kind of user gender prediction method and system | |
CN103870454A (en) | Method and method for recommending data | |
CN107730337A (en) | Information-pushing method and device | |
CN104199872A (en) | Information recommendation method and device | |
CN109977312A (en) | A kind of knowledge base recommender system based on content tab | |
CN107977678A (en) | Method and apparatus for output information | |
CN107864405A (en) | A kind of Forecasting Methodology, device and the computer-readable medium of viewing behavior type | |
CN110413896A (en) | Network information push method, device, computer equipment and storage medium | |
CN112116426A (en) | Method and device for pushing article information | |
CN112100221A (en) | Information recommendation method and device, recommendation server and storage medium | |
CN111626767B (en) | Resource data issuing method, device and equipment | |
Subramanian et al. | UP-GNIV: an expeditious high utility pattern mining algorithm for itemsets with negative utility values | |
CN103870452A (en) | Method and method for recommending data | |
CN109190027A (en) | Multi-source recommended method, terminal, server, computer equipment, readable medium | |
CN114371946B (en) | Information push method and information push server based on cloud computing and big data | |
CN109255682A (en) | A kind of mixed recommendation system towards electronic business system | |
CN111062736A (en) | Model training and clue sequencing method, device and equipment | |
Liu et al. | Social network analysis using big data |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CP03 | Change of name, title or address |
Address after: Room 222, Floor 2, Building 1, Yard A23, North Third Ring West Road, Haidian District, Beijing 100098 Patentee after: Beijing Douyin Information Service Co.,Ltd. Address before: Room 10A, Block B, Yingdu Building, No. 48 Zhichun Road, Haidian District, Beijing 100041 Patentee before: BEIJING BYTEDANCE TECHNOLOGY Co.,Ltd. |
|
CP03 | Change of name, title or address |