CN110019163A - Method, system, equipment and the storage medium of prediction, the recommendation of characteristics of objects - Google Patents
Method, system, equipment and the storage medium of prediction, the recommendation of characteristics of objects Download PDFInfo
- Publication number
- CN110019163A CN110019163A CN201711270287.XA CN201711270287A CN110019163A CN 110019163 A CN110019163 A CN 110019163A CN 201711270287 A CN201711270287 A CN 201711270287A CN 110019163 A CN110019163 A CN 110019163A
- Authority
- CN
- China
- Prior art keywords
- user
- interaction
- feature
- data
- objects
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/21—Design, administration or maintenance of databases
- G06F16/215—Improving data quality; Data cleansing, e.g. de-duplication, removing invalid entries or correcting typographical errors
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- Mathematical Physics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Quality & Reliability (AREA)
- Artificial Intelligence (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of prediction of characteristics of objects, the method for recommendation, system, equipment and storage mediums, the prediction technique includes: to obtain user data and object data, initial characteristics are extracted from the user data and the object data respectively, the initial characteristics include multiple user characteristics and multiple characteristics of objects；The initial characteristics are combined to form multiple interaction features；It is input to multiple perceptron model using the initial characteristics and the interaction feature as training sample, training obtains the prediction model and the interaction coefficient of characteristics of objects；The prediction model is used to predict the characteristics of objects of object, and the interaction coefficient is used to characterize the feature association of the interaction feature.The present invention is based on multiple perceptron model, the initial characteristics and interaction feature of input are trained, the cross correlation between feature can be obtained, take out more valuable feature, while obtaining more preferably prediction model.
Description
Technical field
The present invention relates to machine learning field, a kind of method of special standby prediction, the recommendation for being related to characteristics of objects, is set system
Standby and storage medium.
Background technique
With the development of internet big data and the promotion of computer performance, deep learning is used in respectively more and more
Each industry of row.How the mass data that the user of magnanimity generates, excavate valuable information, show user, it appears increasingly
It is important.
Existing object recommendation system is mainly that analyst expects to determine feature by analyst's strategy after several features
Coefficient provides score calculation formula in conjunction with business experience, calculates score as the score for showing user to sort, still, analyzes
The business experience of teacher is limited, it is impossible to be provided the cross correlation between feature, cannot be taken out more valuable feature.
The object for recommending user as a result, then not necessarily causes the sympathetic response of user well, reduces user experience.
Summary of the invention
The technical problem to be solved by the present invention is to be difficult to provide between feature by analyst in the prior art to overcome
Cross correlation cause the object recommended that cannot meet the defects of user preferences, a kind of prediction of characteristics of objects is provided, is recommended
Method, system, equipment and storage medium.
The present invention is to solve above-mentioned technical problem by following technical proposals:
A kind of prediction technique of characteristics of objects, the prediction technique include:
User data and object data are obtained, is extracted from the user data and the object data respectively initial special
Sign；The initial characteristics include multiple user characteristics and multiple characteristics of objects；
The initial characteristics are combined to form multiple interaction features；
It is input to multiple perceptron model using the initial characteristics and the interaction feature as training sample, training obtains
The interaction coefficient of the prediction model of characteristics of objects and the interaction feature；
The prediction model is used to predict the characteristics of objects of object, and the interaction coefficient is for characterizing the interaction feature
Feature association.
Preferably, the acquisition user data and object data, respectively from the user data and the object data
The step of extracting initial characteristics specifically includes:
Tag analog is carried out to the user data and the object data, and to the user data and mark after tag analog
Object data after label simulation carries out data cleansing；
S11-2, the user data after data cleansing and the object data after data cleansing are normalized；
S11-3, the initial characteristics are extracted from the user data after normalization and the object data after normalization.
It the initial characteristics is combined the step of forming multiple interaction features specifically includes preferably, described:
The initial characteristics are subjected to traversal combination or self-defined combination forms multiple interaction features.
Preferably, described be input to multi-layer perception (MLP) mould using the initial characteristics and the interaction feature as training sample
The step of interaction coefficient of type, the trained prediction model for obtaining characteristics of objects and the interaction feature, specifically includes:
Define the initial model parameter of multiple perceptron model；Model parameter includes the initial coefficients and friendship of initial characteristics
The initial interaction coefficient of mutual feature；
The initial characteristics and interaction feature that the extraction of the first acquisition time is acquired from the training sample input the multilayer
Perceptron model；
The characteristics of objects that the multiple perceptron model exports is extracted with from the training sample using loss function
The second acquisition time obtain characteristics of objects be compared to obtain the loss of the multiple perceptron model, and adjust automatically institute
State the model parameter of multiple perceptron model；Second acquisition time is later than first acquisition time；
Iteration executes the step of loss for calculating the multiple perceptron model, until the loss is in first threshold range
It is interior, obtain the prediction model and the interaction coefficient；And/or iteration executes the loss for calculating the multiple perceptron model
The step of, until the number of iterations reaches preset maximum number of iterations, obtain the prediction model and the interaction coefficient.
Preferably, the loss function is cross entropy loss function.
Preferably, the user characteristics comprise at least one of the following:
User's gender, user gradation, user device address, user network information, user member's rank, user's purchasing power and
Predilection grade；
Characteristics of objects comprises at least one of the following:
The click volume of object, sharing amount, pageview, light exposure, GMV (gross turnover), exposes clicking rate, is right at the amount of thumbing up
Statistic, object quality point, object author's mass point, object gender and object scale of price as entering the detailed page of quotient.
A kind of electronic equipment including memory, processor and stores the meter that can be run on a memory and on a processor
Calculation machine program, the processor realize the pre- of the characteristics of objects of above-mentioned each preference any combination when executing the computer program
Survey method.
A kind of computer readable storage medium, is stored thereon with computer program, and the computer program is held by processor
The step of prediction technique of the characteristics of objects of above-mentioned each preference any combination is realized when row.
A kind of object recommendation method, the object recommendation method the following steps are included:
It responds user and accesses object recommendation system, the candidate target for intending recommending to the user is determined according to user information；
Prediction model is obtained using above-mentioned prediction technique；
Characteristics of objects is predicted according to the prediction model and calculates the scoring of each candidate target；
The candidate target is ranked up according to the sequence of scoring from high to low.
Preferably, the object recommendation method further include:
Recommendation score is greater than the candidate target of scoring threshold value.
A kind of electronic equipment including memory, processor and stores the meter that can be run on a memory and on a processor
Calculation machine program, the processor realize the object recommendation side of above-mentioned each preference any combination when executing the computer program
Method.
A kind of computer readable storage medium, is stored thereon with computer program, and the computer program is held by processor
The step of object recommendation method of above-mentioned each preference any combination is realized when row.
A kind of forecasting system of characteristics of objects, the forecasting system include:
Feature obtains module, for obtaining user data and object data and from the user data and the object data
Middle extraction initial characteristics, the initial characteristics include multiple user characteristics and multiple characteristics of objects；
Feature interaction module, for being combined to form multiple interaction features by the initial characteristics；
Training module, for being input to multi-layer perception (MLP) using the initial characteristics and the interaction feature as training sample
Model, training obtain the prediction model and the interaction coefficient of characteristics of objects；
The prediction model is used to predict the characteristics of objects of object, and the interaction coefficient is for characterizing the interaction feature
Feature association.
Preferably, the feature obtain module include tag analog unit, data cleansing unit, normalized unit and
Feature extraction unit；
The tag analog unit is used to carry out tag analog to the user data and the object data；
The data cleansing unit is used to carry out the object data after the user data and tag analog after tag analog
Data cleansing；
The normalized unit be used for the user data after data cleansing and the object data after data cleansing into
Row normalized；
The feature extraction unit from the user data after normalization and the object data after normalization for extracting institute
State initial characteristics.
Preferably, the feature interaction module is used to the initial characteristics carrying out traversal combination or self-defined combination is formed
Multiple interaction features.
Preferably, the training module includes:
Parameter initialization unit, for defining the initial model parameter of multiple perceptron model；Model parameter includes just
The initial coefficients of beginning feature and the initial interaction coefficient of interaction feature；
Input unit, for acquiring the initial characteristics and interaction feature of the extraction of the first acquisition time from the training sample
Input the multiple perceptron model；
Comparing unit, characteristics of objects for being exported the multiple perceptron model using loss function with from the instruction
Practice the characteristics of objects that the second acquisition time extracted in sample obtains to be compared to obtain the loss of the multiple perceptron model,
And the model parameter of multiple perceptron model described in adjust automatically；Second acquisition time is later than first acquisition time；
Call unit calls the input unit and the comparing unit for repeating, until the loss is in the first threshold
It is worth in range, obtains the prediction model and the interaction coefficient；And/or it repeats to call the input unit and the comparison
Unit obtains the prediction model and the interaction coefficient until the number of iterations reaches preset maximum number of iterations.
Preferably, the loss function is cross entropy loss function.
Preferably, the user characteristics comprise at least one of the following:
User's gender, user gradation, user device address, user network information, user member's rank, user's purchasing power and
Predilection grade；
Characteristics of objects comprises at least one of the following:
The click volume of object, the amount of thumbing up, sharing amount, pageview, light exposure, GMV, exposure clicking rate, that object enters quotient is detailed
Statistic, object quality point, object author's mass point, object gender and the object scale of price of page.
A kind of object recommendation system, the object recommendation system include candidate target determining module, grading module, sequence mould
Block and the prediction model obtained using above-mentioned forecasting system；
Candidate target determining module accesses object recommendation system for responding user, is intended according to user information determination to institute
State the candidate target of user's recommendation；
Grading module is used to predict characteristics of objects according to the prediction model and calculates the scoring of each candidate target；
Sorting module is used to for the candidate target being ranked up according to the sequence of scoring from high to low.
Preferably, the object recommendation system further includes recommending module；
The recommending module is greater than the candidate target of scoring threshold value for recommendation score.
The positive effect of the present invention is that: the present invention is based on multiple perceptron model, initial characteristics to input and
Interaction feature is trained, and can obtain the cross correlation between feature, takes out more valuable feature, while obtaining more
Excellent prediction model.
Detailed description of the invention
Fig. 1 is a kind of flow chart of the prediction technique of characteristics of objects of the embodiment of the present invention 1.
Fig. 2 is the specific flow chart of step 101 in the embodiment of the present invention 1.
Fig. 3 is a kind of specific flow chart of implementation of step 103 in the embodiment of the present invention 1.
Fig. 4 is the specific flow chart of another implementation of step 103 in the embodiment of the present invention 1.
Fig. 5 is the hardware structural diagram of a kind of electronic equipment of the embodiment of the present invention 2.
Fig. 6 is a kind of flow chart of object recommendation method of the embodiment of the present invention 4.
Fig. 7 is a kind of structural block diagram of the forecasting system of characteristics of objects of the embodiment of the present invention 7.
Fig. 8 is a kind of structural block diagram of object recommendation system of the embodiment of the present invention 8.
Specific embodiment
The present invention is further illustrated below by the mode of embodiment, but does not therefore limit the present invention to the reality
It applies among a range.
Embodiment 1
As shown in Figs 1-4, a kind of prediction technique of characteristics of objects, comprising:
Step 101 obtains user data and object data, extracts from user data and object data respectively initial special
Sign；Initial characteristics include multiple user characteristics and multiple characteristics of objects；Step 102, initial characteristics are combined to be formed it is multiple
Interaction feature；
Step 103 is input to multiple perceptron model using initial characteristics and interaction feature as training sample, and training obtains
The prediction model of characteristics of objects and the interaction coefficient of interaction feature；
Prediction model is used to predict the characteristics of objects of object, and interaction coefficient is used to characterize the feature association of interaction feature.
It needs, the object in the present invention can be books, article (such as Internet news, web documents) and commodity
Deng.The content information of object information namely books, the content information of article, the recommended information etc. of commodity.
In step 101, can choose at least one of following characteristics be used as user characteristics: user's gender, user gradation,
User device address, user network information, user member's rank, user's purchasing power and predilection grade etc.；And selection is following special
At least one of sign is as characteristics of objects: the click volume of object, the amount of thumbing up, sharing amount, pageview, light exposure, GMV, exposure
Clicking rate, object enter the statistic of the detailed page of quotient, object quality point, object author's mass point, object gender and object price etc.
Grade etc..
In the present embodiment, as shown in Fig. 2, step 101 specifically includes:
Step 101-1, tag analog is carried out to user data and object data；
Wherein, multi-layer perception (MLP) is a kind of machine learning algorithm for having supervision, training pattern characteristics of needs data, it is also desirable to
Label data, it is critically important a part that how the data in training set, which label,.Thinking of the invention is: showing from client
To the material in the exposure log of the object recommendation system of user, filtering out user's click, labelling is 1, remaining material is beaten
Label 0, this process need to be associated with exposure log sheet and click logs table.
Step 101-2, data cleansing is carried out to the object data after the user data and tag analog after tag analog；
The purpose of data cleansing is processing exceptional sample, has following three kinds to the processing mode of sample in the present invention:
(a) full line sample data is directly abandoned, such as when label filtration, since the number that label is 0 is much larger than mark
Label are 1 numbers, need the data random screening by a certain percentage for being 0 to label；
(b) absent field is filled with default value, for example handles user's sex character in the present invention, when taking less than user's property
When other, default value -1 is exactly mended, is equivalent to as a kind of independent classification；
(c) missing data is filled with to the average value of permutation data, this way, which is equivalent to, weakens missing item to other
The influence of feature.
Step 101-3, the user data after data cleansing and the object data after data cleansing are normalized；
Wherein, the purpose of normalized is so that data dimension is consistent, and the data for keeping script in disorder are easier by algorithm
Processing, that is, the format that data preparation is needed at deep learning, what is generated after processing is the feature mark that machine learning needs
Sign the libsvm format of composition
Step 101-4, initial characteristics are extracted from the user data after normalization and the object data after normalization.
It is specifically included in step 102:
Initial characteristics are subjected to traversal combination or self-defined combination forms multiple interaction features.
Wherein, after having selected initial characteristics, the feature that can also be interacted according to the customized selected needs of service feature,
It is trained to obtain the interaction coefficient after interaction coefficient is further trained by multiple perceptron model again, then determines that monitoring is special
Relevance between sign.
In the present embodiment, as shown in figure 3, step 103 specifically includes:
Step 103-1, the initial model parameter of multiple perceptron model is defined；Model parameter includes the first of initial characteristics
The initial interaction coefficient of beginning coefficient and interaction feature；
Wherein, model parameter is initialized, generallys use the method to stochastic parameter assignment, it can also be each parameter
Initial value be both configured to 0, initialized in the present invention using gaussian random assignment；
Step 103-2, initial characteristics and the interaction feature input that the extraction of the first acquisition time is acquired from training sample are more
Layer perceptron model；
Step 103-3, the characteristics of objects that multiple perceptron model exports is mentioned with from training sample using loss function
The characteristics of objects that the second acquisition time taken obtains is compared to obtain the loss of multiple perceptron model, and adjust automatically multilayer
The model parameter of perceptron model；Second acquisition time is later than the first acquisition time；Loss function, which can be used, intersects entropy loss letter
Number.
Wherein, the loss of calculating be one can portray model final step obtain output with the phase from training set
Hope the generality index of gap between output, there are many loss functions, and the present invention uses cross entropy loss function；Align model ginseng
Several processes is the process of a self study, and after loss function, aiming at for study is changed by a large amount of training steps
It is apt to the value of each parameter, so that the present invention uses gradient descent algorithm by minimization of loss.
Step 103-4, iteration executes step 103-2 and step 103-3, until loss obtains in first threshold range
Prediction model and interaction coefficient；
Wherein, when penalty values reach the standard of setting, prediction model can be obtained as optimal prediction with deconditioning
Model；
As shown in figure 4, providing the another kind of step 103-4 ' if penalty values in the training process are not up to always standard
Embodiment:
Step 103-4 ', iteration execute step 103-2 and step 103-3, until the number of iterations reaches preset maximum and changes
Generation number, obtains prediction model and interaction coefficient.
Wherein, once repetitive exercise number reaches highest number, such as 100 times, loss reduction in 100 times is selected at this time
The model once exported is optimum prediction model, it should be noted that as long as above two iterative manner has one kind to reach setting
Standard stop iteration, and obtain prediction model.
It should be noted that the interaction coefficient that the present embodiment obtains can be used to characterize the cross correlation of interaction feature,
For example, illustrating to have no relevance between interaction feature, being mutually indepedent between interaction feature if obtained interaction coefficient is 0
It has no effect on；If obtained interaction coefficient is not 0, illustrate the relevance for having implicit between interaction feature, is evaluating or pushing away
During recommending object, then the influence for the implicit relevance for further considering the interaction feature is needed.
In the present embodiment, it is based on multiple perceptron model, the initial characteristics and interaction feature of input are trained, it can
The cross correlation between feature is obtained, more valuable feature is taken out, while obtaining more preferably prediction model.
Embodiment 2
A kind of electronic equipment including memory, processor and stores the meter that can be run on a memory and on a processor
Calculation machine program, the processor realize the prediction technique of the characteristics of objects in embodiment 1 when executing the computer program.
Fig. 5 is the structural schematic diagram for a kind of electronic equipment that the embodiment of the present invention 2 provides.Fig. 5, which is shown, to be suitable for being used in fact
The block diagram of the example electronic device 30 of existing embodiment of the present invention.The electronic equipment 30 that Fig. 5 is shown is only an example, no
The function and use scope for coping with the embodiment of the present invention bring any restrictions.
As shown in figure 5, electronic equipment 30 can be showed in the form of universal computing device, such as it can set for server
It is standby.The component of electronic equipment 30 can include but is not limited to: at least one processor 31, at least one processor 32, connection are not
The bus 33 of homologous ray component (including memory 32 and processor 31).
Bus 33 includes data/address bus, address bus and control bus.
Memory 32 may include volatile memory, such as random access memory (RAM) 321 and/or cache
Memory 322 can further include read-only memory (ROM) 323.
Memory 32 can also include the program means 325 with one group of (at least one) program module 324, such journey
Sequence module 324 includes but is not limited to: operating system, one or more application program, other program modules and program data,
It may include the realization of network environment in each of these examples or certain combination.
Processor 31 by operation storage computer program in memory 32, thereby executing various function application and
Data processing, such as the storage method of object information provided by the embodiment of the present invention 1.
Electronic equipment 30 can also be communicated with one or more external equipments 34 (such as keyboard, sensing equipment etc.).It is this
Communication can be carried out by input/output (I/O) interface 35.Also, electronic equipment 30 can also by network adapter 36 with
One or more network (such as local area network (LAN), wide area network (WAN) and/or public network, such as internet) communication.Net
Network adapter 36 is communicated by bus 33 with other modules of electronic equipment 30.It should be understood that although not shown in the drawings, can tie
It closes electronic equipment 30 and uses other hardware and/or software module, including but not limited to: microcode, device driver, redundancy processing
Device, external disk drive array, RAID (disk array) system, tape drive and data backup storage system etc..
It should be noted that although being referred to several units/modules or subelement/mould of electronic equipment in the above detailed description
Block, but it is this division be only exemplary it is not enforceable.In fact, being retouched above according to presently filed embodiment
The feature and function for two or more units/modules stated can embody in a units/modules.Conversely, above description
A units/modules feature and function can with further division be embodied by multiple units/modules.
Embodiment 3
A kind of computer readable storage medium, is stored thereon with computer program, and the computer program is held by processor
The step of prediction of the characteristics of objects in embodiment 1 is realized when row.
Wherein, what readable storage medium storing program for executing can use more specifically can include but is not limited to: portable disc, hard disk, random
Access memory, read-only memory, erasable programmable read only memory, light storage device, magnetic memory device or above-mentioned times
The suitable combination of meaning.
In possible embodiment, the present invention is also implemented as a kind of form of program product comprising program generation
Code, when described program product is run on the terminal device, said program code is realized in fact for executing the terminal device
Apply the step in the risk evaluating method in example 1.
Wherein it is possible to be write with any combination of one or more programming languages for executing program of the invention
Code, said program code can be executed fully on a user device, partly execute on a user device, is only as one
Vertical software package executes, part executes on a remote device or executes on a remote device completely on a user device for part.
Embodiment 4
As shown in fig. 6, a kind of object recommendation method, comprising the following steps:
Step 201, response user access object recommendation system, and it is right to intend candidate recommended to the user according to user information determination
As；
Step 202 obtains prediction model；
Specifically, obtaining prediction model using the prediction technique in embodiment 1.
Step 203 predicts characteristics of objects according to prediction model and calculates the scoring of each candidate target；
Specifically, the weight of each object feature of settable prediction model output, passes through the value and weight of each feature
Calculate the scoring of each object.
Candidate target is ranked up by step 204 according to the sequence of scoring from high to low.
Object is ranked up according to the sequence of scoring from high to low, realizes and does personalized prediction to the preference of user,
So that recommendation effect is more preferably.
In the present embodiment, after step 204 further include:
Step 205, recommendation score are greater than the candidate target of scoring threshold value.
In the present embodiment, prediction model can accurately predict the characteristics of objects of each object, be obtained according to the prediction model
Characteristics of objects value it is more accurate, so as to quantify the scoring of each object by the value of computing object feature, and according to scoring
Object recommendation is carried out, accuracy greatly improves.
Embodiment 5
A kind of equipment of risk control, including memory, processor and storage can transport on a memory and on a processor
Capable computer program, the processor realize the object recommendation method in embodiment 4 when executing the computer program.
Embodiment 6
A kind of computer readable storage medium, is stored thereon with computer program, and the computer program is held by processor
The step of object recommendation method in embodiment 4 is realized when row.
Embodiment 7
As shown in fig. 7, a kind of forecasting system of characteristics of objects, including feature obtain module 11,12 and of feature interaction module
Training module 13；
Feature obtains module 11 for obtaining user data and object data and extracting from user data and object data
Initial characteristics, initial characteristics include multiple user characteristics and multiple characteristics of objects；
Wherein it is possible to select at least one of following characteristics as user characteristics: user's gender, user gradation, user
Device address, user network information, user member's rank, user's purchasing power and predilection grade etc.；And in selection following characteristics
At least one as characteristics of objects: the click volume of object, the amount of thumbing up, sharing amount, pageview, light exposure, GMV, exposure are clicked
Rate, object enter the statistic of the detailed page of quotient, object quality point, object author's mass point, object gender and object scale of price etc..
Initial characteristics for being combined to form multiple interaction features by feature interaction module 12；
Wherein, feature interaction module 12 is also used to carry out initial characteristics traversal combination or self-defined combination forms multiple friendships
Mutual feature.After having selected initial characteristics, the feature that can also be interacted according to the customized selected needs of service feature, then by more
Layer perceptron model is trained to obtain the interaction coefficient after interaction coefficient is further trained, then determines between monitored characteristic
Relevance.
Training module 13 is used to be input to multiple perceptron model using initial characteristics and interaction feature as training sample, instructs
Get the prediction model of characteristics of objects and the interaction coefficient of interaction feature；
Prediction model is used to predict the characteristics of objects of object, and interaction coefficient is used to characterize the feature association of interaction feature.
In the present embodiment, feature obtains module 11 and specifically includes tag analog unit 111, data cleansing unit 112, normalizing
Change processing unit 113 and feature extraction unit 114；
Tag analog unit 111 is used to carry out tag analog to user data and object data；
Wherein, multi-layer perception (MLP) is a kind of machine learning algorithm for having supervision, training pattern characteristics of needs data, it is also desirable to
Label data, it is critically important a part that how the data in training set, which label,.Thinking of the invention is: showing from client
To the material in the exposure log of the object recommendation system of user, filtering out user's click, labelling is 1, remaining material is beaten
Label 0, this process need to be associated with exposure log sheet and click logs table.
Data cleansing unit 112 is for counting the object data after the user data and tag analog after tag analog
According to cleaning；
The purpose of data cleansing is processing exceptional sample, has following three kinds to the processing mode of sample in the present invention:
(a) full line sample data is directly abandoned, such as when label filtration, since the number that label is 0 is much larger than mark
Label are 1 numbers, need the data random screening by a certain percentage for being 0 to label；
(b) absent field is filled with default value, for example handles user's sex character in the present invention, when taking less than user's property
When other, default value -1 is exactly mended, is equivalent to as a kind of independent classification；
(c) missing data is filled with to the average value of permutation data, this way, which is equivalent to, weakens missing item to other
The influence of feature.
Normalized unit 113 is carried out for the object data to the user data after data cleansing and after data cleansing
Normalized；
Wherein, the purpose of normalized is so that data dimension is consistent, and the data for keeping script in disorder are easier by algorithm
Processing, that is, the format that data preparation is needed at deep learning, what is generated after processing is the feature mark that machine learning needs
Sign the libsvm format of composition
Feature extraction unit 114 is used to extract from the user data after normalization and the object data after normalization initial
Feature.
In the present embodiment, training module 13 specifically includes parameter initialization unit 131, input unit 132, comparing unit
133 and call unit 134:
Parameter initialization unit 131 is used to define the initial model parameter of multiple perceptron model；Model parameter includes
The initial coefficients of initial characteristics and the initial interaction coefficient of interaction feature；
Wherein, model parameter is initialized, generallys use the method to stochastic parameter assignment, it can also be each parameter
Initial value be both configured to 0, initialized in the present invention using gaussian random assignment；
Input unit 132 is used to acquire the initial characteristics that the first acquisition time extracts from training sample and interaction feature is defeated
Enter multiple perceptron model；
Characteristics of objects that comparing unit 133 is used to export multiple perceptron model using loss function and from training sample
The characteristics of objects that second acquisition time of middle extraction obtains is compared to obtain the loss of multiple perceptron model, and adjust automatically
The model parameter of multiple perceptron model；Second acquisition time is later than the first acquisition time；Cross entropy damage can be used in loss function
Lose function.
Wherein, the loss of calculating be one can portray model final step obtain output with the phase from training set
Hope the generality index of gap between output, there are many loss functions, and the present invention uses cross entropy loss function；Align model ginseng
Several processes is the process of a self study, and after loss function, aiming at for study is changed by a large amount of training steps
It is apt to the value of each parameter, so that the present invention uses gradient descent algorithm by minimization of loss.
Call unit 134 calls input unit 132 and comparing unit 133 for repeating, until loss is in first threshold model
In enclosing, prediction model and interaction coefficient are obtained；And/or call unit 134 repeats to call input unit 132 and comparing unit
133, until the number of iterations reaches preset maximum number of iterations, obtain prediction model and interaction coefficient.
Wherein, when penalty values reach the standard of setting, prediction model can be obtained as optimal prediction with deconditioning
Model, if penalty values in the training process are not up to always standard, once repetitive exercise number reaches highest number, such as
100 times, select the model of loss reduction in 100 times once exported for optimum prediction model at this time.Wherein, the interaction system obtained
Number can be used to characterize the cross correlation of interaction feature, for example, if obtained interaction coefficient is 0, illustrate interaction feature it
Between have no relevance, had no effect on independently of each other between interaction feature；If obtained interaction coefficient is not 0, illustrate to hand over
There is implicit relevance between mutual feature, during evaluation or recommended, then needs further to consider the interaction feature
The influence of implicit relevance.
In the present embodiment, it is based on multiple perceptron model, the initial characteristics and interaction feature of input are trained, it can
The cross correlation between feature is obtained, more valuable feature is taken out, while obtaining more preferably prediction model.
Embodiment 8
As shown in figure 8, a kind of object recommendation system, including candidate target determining module 21, grading module 22, sorting module
23 and prediction model 14；Prediction model 14 is obtained using the forecasting system in embodiment 7.
Candidate target determining module 21 for respond user access object recommendation system, according to user information determine intend to
The candidate target that family is recommended；
Grading module 22 is used to predict characteristics of objects according to prediction model and calculates the scoring of each candidate target；Specifically
, the weight of each object feature of settable prediction model output passes through each object of value and weight calculation of each feature
Scoring.
Sorting module 23 is used to for candidate target being ranked up according to the sequence of scoring from high to low.By object according to scoring
Sequence from high to low is ranked up, and is realized and is done personalized prediction to the preference of user, so that recommendation effect is more preferably.
If can directly export object score, this implementation by the prediction model that the forecasting system in embodiment 7 obtains
Object recommendation system in example is then not necessarily to grading module, sorting module directly according to the object score of prediction model output directly into
Row object order.
Object recommendation system further includes recommending module 24, and the candidate target of scoring threshold value is greater than for recommendation score.
In the present embodiment, prediction model can accurately predict the characteristics of objects of each object, be obtained according to the prediction model
Characteristics of objects value it is more accurate, so as to quantify the scoring of each object by the value of computing object feature, and according to scoring
Object recommendation is carried out, accuracy greatly improves.
Although specific embodiments of the present invention have been described above, it will be appreciated by those of skill in the art that this is only
For example, protection scope of the present invention is to be defined by the appended claims.Those skilled in the art without departing substantially from
Under the premise of the principle and substance of the present invention, many changes and modifications may be made, but these change and
Modification each falls within protection scope of the present invention.
Claims (20)
1. a kind of prediction technique of characteristics of objects, which is characterized in that the prediction technique includes:
User data and object data are obtained, extracts initial characteristics from the user data and the object data respectively；Institute
Stating initial characteristics includes multiple user characteristics and multiple characteristics of objects；
The initial characteristics are combined to form multiple interaction features；
It is input to multiple perceptron model using the initial characteristics and the interaction feature as training sample, training obtains object
The interaction coefficient of the prediction model of feature and the interaction feature；
The prediction model is used to predict the characteristics of objects of object, and the interaction coefficient is used to characterize the feature of the interaction feature
Relevance.
2. prediction technique as described in claim 1, which is characterized in that the acquisition user data and object data, respectively from
The step of initial characteristics are extracted in the user data and the object data specifically includes:
Tag analog is carried out to the user data and the object data, and to the user data and label mould after tag analog
Object data after quasi- carries out data cleansing；
User data after data cleansing and the object data after data cleansing are normalized；
The initial characteristics are extracted from the user data after normalization and the object data after normalization.
3. prediction technique as described in claim 1, which is characterized in that it is described the initial characteristics are combined to be formed it is multiple
The step of interaction feature, specifically includes:
The initial characteristics are subjected to traversal combination or self-defined combination forms multiple interaction features.
4. prediction technique as described in claim 1, which is characterized in that described to make the initial characteristics and the interaction feature
It is input to multiple perceptron model for training sample, training obtains the interaction of the prediction model and the interaction feature of characteristics of objects
The step of coefficient, specifically includes:
Define the initial model parameter of multiple perceptron model；Model parameter includes initial coefficients and the interaction spy of initial characteristics
The initial interaction coefficient of sign；
The initial characteristics and interaction feature that the extraction of the first acquisition time is acquired from the training sample input the Multilayer Perception
Machine model；
Using loss function by characteristics of objects that the multiple perceptron model exports with extracted from the training sample the
The characteristics of objects that two acquisition times obtain is compared to obtain the loss of the multiple perceptron model, and more described in adjust automatically
The model parameter of layer perceptron model；Second acquisition time is later than first acquisition time；
Iteration executes the step of loss for calculating the multiple perceptron model, until described lose in first threshold range,
Obtain the prediction model and the interaction coefficient；And/or iteration executes the step for calculating the loss of the multiple perceptron model
Suddenly, until the number of iterations reaches preset maximum number of iterations, the prediction model and the interaction coefficient are obtained.
5. prediction technique as claimed in claim 4, which is characterized in that the loss function is cross entropy loss function.
6. prediction technique as described in claim 1, which is characterized in that the user characteristics comprise at least one of the following:
User's gender, user gradation, user device address, user network information, user member's rank, user's purchasing power and preference
Scoring；
Characteristics of objects comprises at least one of the following:
The click volume of object, the amount of thumbing up, sharing amount, pageview, light exposure, GMV, exposure clicking rate, object enter the detailed page of quotient
Statistic, object quality point, object author's mass point, object gender and object scale of price.
7. a kind of electronic equipment including memory, processor and stores the calculating that can be run on a memory and on a processor
Machine program, which is characterized in that the processor is realized as claimed in any one of claims 1 to 6 when executing the computer program
The prediction technique of characteristics of objects.
8. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer program quilt
The step of processor realizes the prediction technique of characteristics of objects as claimed in any one of claims 1 to 6 when executing.
9. a kind of object recommendation method, which is characterized in that the object recommendation method the following steps are included:
It responds user and accesses object recommendation system, the candidate target for intending recommending to the user is determined according to user information；
Prediction model is obtained using prediction technique described in any one of claim 1-6；
Characteristics of objects is predicted according to the prediction model and calculates the scoring of each candidate target；
The candidate target is ranked up according to the sequence of scoring from high to low.
10. object recommendation method as claimed in claim 9, which is characterized in that the object recommendation method further include:
Recommendation score is greater than the candidate target of scoring threshold value.
11. a kind of electronic equipment including memory, processor and stores the calculating that can be run on a memory and on a processor
Machine program, which is characterized in that realize that object described in claim 9 or 10 pushes away when the processor executes the computer program
Recommend method.
12. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer program
The step of object recommendation method described in claim 9 or 10 is realized when being executed by processor.
13. a kind of forecasting system of characteristics of objects, which is characterized in that the forecasting system includes:
Feature obtains module, for obtaining user data and object data and mentioning from the user data and the object data
Initial characteristics are taken, the initial characteristics include multiple user characteristics and multiple characteristics of objects；
Feature interaction module, for being combined to form multiple interaction features by the initial characteristics；
Training module, for being input to multi-layer perception (MLP) mould using the initial characteristics and the interaction feature as training sample
Type, training obtain the prediction model of characteristics of objects and the interaction coefficient of the interaction feature；
The prediction model is used to predict the characteristics of objects of object, and the interaction coefficient is used to characterize the feature of the interaction feature
Relevance.
14. forecasting system as claimed in claim 13, which is characterized in that it includes tag analog list that the feature, which obtains module,
Member, data cleansing unit, normalized unit and feature extraction unit；
The tag analog unit is used to carry out tag analog to the user data and the object data；
The data cleansing unit is used to carry out data to the object data after the user data and tag analog after tag analog
Cleaning；
The normalized unit is for returning the user data after data cleansing and the object data after data cleansing
One change processing；
The feature extraction unit is described first for extracting from the user data after normalization and the object data after normalization
Beginning feature.
15. forecasting system as claimed in claim 13, which is characterized in that the feature interaction module is used for the initial spy
Sign carries out traversal combination or self-defined combination forms multiple interaction features.
16. forecasting system as claimed in claim 13, which is characterized in that the training module includes:
Parameter initialization unit, for defining the initial model parameter of multiple perceptron model；Model parameter includes initial special
The initial coefficients of sign and the initial interaction coefficient of interaction feature；
Input unit, for acquiring initial characteristics and the interaction feature input of the extraction of the first acquisition time from the training sample
The multiple perceptron model；
Comparing unit, characteristics of objects for being exported the multiple perceptron model using loss function with from the trained sample
The characteristics of objects that the second acquisition time extracted in this obtains is compared to obtain the loss of the multiple perceptron model, and certainly
The model parameter of the dynamic adjustment multiple perceptron model；Second acquisition time is later than first acquisition time；
Call unit calls the input unit and the comparing unit for repeating, until the loss is in first threshold model
In enclosing, the prediction model and the interaction coefficient are obtained；And/or repeat to call the input unit and the comparing unit,
Until the number of iterations reaches preset maximum number of iterations, the prediction model and the interaction coefficient are obtained.
17. forecasting system as claimed in claim 16, which is characterized in that the loss function is cross entropy loss function.
18. forecasting system as claimed in claim 13, which is characterized in that the user characteristics comprise at least one of the following:
User's gender, user gradation, user device address, user network information, user member's rank, user's purchasing power and preference
Scoring；
Characteristics of objects comprises at least one of the following:
The click volume of object, the amount of thumbing up, sharing amount, pageview, light exposure, GMV, exposure clicking rate, object enter the detailed page of quotient
Statistic, object quality point, object author's mass point, object gender and object scale of price.
19. a kind of object recommendation system, which is characterized in that the object recommendation system includes candidate target determining module, scoring
Module, sorting module and the prediction model obtained using forecasting system described in any one of claim 13-18；
Candidate target determining module accesses object recommendation system for responding user, is intended according to user information determination to the use
The candidate target that family is recommended；
Grading module is used to predict characteristics of objects according to the prediction model and calculates the scoring of each candidate target；
Sorting module is used to for the candidate target being ranked up according to the sequence of scoring from high to low.
20. object recommendation system as claimed in claim 19, which is characterized in that the object recommendation system further includes recommending mould
Block；
The recommending module is greater than the candidate target of scoring threshold value for recommendation score.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711270287.XA CN110019163A (en) | 2017-12-05 | 2017-12-05 | Method, system, equipment and the storage medium of prediction, the recommendation of characteristics of objects |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711270287.XA CN110019163A (en) | 2017-12-05 | 2017-12-05 | Method, system, equipment and the storage medium of prediction, the recommendation of characteristics of objects |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110019163A true CN110019163A (en) | 2019-07-16 |
Family
ID=67186010
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711270287.XA Pending CN110019163A (en) | 2017-12-05 | 2017-12-05 | Method, system, equipment and the storage medium of prediction, the recommendation of characteristics of objects |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110019163A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110442804A (en) * | 2019-08-13 | 2019-11-12 | 北京市商汤科技开发有限公司 | A kind of training method, device, equipment and the storage medium of object recommendation network |
| CN110765363A (en) * | 2019-09-27 | 2020-02-07 | 复旦大学 | Depth recommendation model based on Gaussian distribution representation |
| CN111242752A (en) * | 2020-04-24 | 2020-06-05 | 支付宝(杭州)信息技术有限公司 | Method and system for determining recommended object based on multi-task prediction |
| CN111382190A (en) * | 2020-03-04 | 2020-07-07 | 腾讯科技（深圳）有限公司 | Object recommendation method and device based on intelligence and storage medium |
| CN111680221A (en) * | 2020-08-11 | 2020-09-18 | 腾讯科技（深圳）有限公司 | Information recommendation method, device, equipment and computer readable storage medium |
| CN112989202A (en) * | 2021-04-02 | 2021-06-18 | 常熟理工学院 | Personalized recommendation method and system based on dynamic network embedding |
| CN113139856A (en) * | 2021-05-14 | 2021-07-20 | 青岛聚看云科技有限公司 | Movie and television member package recommendation method and device |
| WO2021208535A1 (en) * | 2020-04-14 | 2021-10-21 | 华为技术有限公司 | Recommendation method and device based on automatic feature grouping |
| CN113744009A (en) * | 2020-05-29 | 2021-12-03 | 北京沃东天骏信息技术有限公司 | Target object output method and device, computer readable medium and electronic equipment |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102346899A (en) * | 2011-10-08 | 2012-02-08 | 亿赞普(北京)科技有限公司 | Method and device for predicting advertisement click rate based on user behaviors |
| CN104462593A (en) * | 2014-12-29 | 2015-03-25 | 北京奇虎科技有限公司 | Method and device for providing user personalized resource message pushing |
| CN105160548A (en) * | 2015-08-20 | 2015-12-16 | 北京奇虎科技有限公司 | Method and apparatus for predicting advertisement click-through rate |
| CN105975472A (en) * | 2015-12-09 | 2016-09-28 | 乐视网信息技术（北京）股份有限公司 | Method and device for recommendation |
| US20170103451A1 (en) * | 2015-10-12 | 2017-04-13 | Yandex Europe Ag | Method and system of determining an optimal value of an auction parameter for a digital object |
| CN107220217A (en) * | 2017-05-31 | 2017-09-29 | 北京京东尚科信息技术有限公司 | Characteristic coefficient training method and device that logic-based is returned |
| CN107316234A (en) * | 2017-07-21 | 2017-11-03 | 北京京东尚科信息技术有限公司 | Personalized commercial Forecasting Methodology and device |
| CN107423442A (en) * | 2017-08-07 | 2017-12-01 | 火烈鸟网络(广州)股份有限公司 | Method and system, storage medium and computer equipment are recommended in application based on user's portrait behavioural analysis |
-
2017
- 2017-12-05 CN CN201711270287.XA patent/CN110019163A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102346899A (en) * | 2011-10-08 | 2012-02-08 | 亿赞普(北京)科技有限公司 | Method and device for predicting advertisement click rate based on user behaviors |
| CN104462593A (en) * | 2014-12-29 | 2015-03-25 | 北京奇虎科技有限公司 | Method and device for providing user personalized resource message pushing |
| CN105160548A (en) * | 2015-08-20 | 2015-12-16 | 北京奇虎科技有限公司 | Method and apparatus for predicting advertisement click-through rate |
| US20170103451A1 (en) * | 2015-10-12 | 2017-04-13 | Yandex Europe Ag | Method and system of determining an optimal value of an auction parameter for a digital object |
| CN105975472A (en) * | 2015-12-09 | 2016-09-28 | 乐视网信息技术（北京）股份有限公司 | Method and device for recommendation |
| CN107220217A (en) * | 2017-05-31 | 2017-09-29 | 北京京东尚科信息技术有限公司 | Characteristic coefficient training method and device that logic-based is returned |
| CN107316234A (en) * | 2017-07-21 | 2017-11-03 | 北京京东尚科信息技术有限公司 | Personalized commercial Forecasting Methodology and device |
| CN107423442A (en) * | 2017-08-07 | 2017-12-01 | 火烈鸟网络(广州)股份有限公司 | Method and system, storage medium and computer equipment are recommended in application based on user's portrait behavioural analysis |
Non-Patent Citations (2)
| Title |
|---|
| 田嫦丽: "互联网广告点击率预估模型中特征", 《计算机应用研究》 * |
| 程显毅: "《深度学习与R语言》", 30 June 2016, 机械工业出版社 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110442804A (en) * | 2019-08-13 | 2019-11-12 | 北京市商汤科技开发有限公司 | A kind of training method, device, equipment and the storage medium of object recommendation network |
| CN110765363A (en) * | 2019-09-27 | 2020-02-07 | 复旦大学 | Depth recommendation model based on Gaussian distribution representation |
| CN111382190A (en) * | 2020-03-04 | 2020-07-07 | 腾讯科技（深圳）有限公司 | Object recommendation method and device based on intelligence and storage medium |
| WO2021208535A1 (en) * | 2020-04-14 | 2021-10-21 | 华为技术有限公司 | Recommendation method and device based on automatic feature grouping |
| CN113536097A (en) * | 2020-04-14 | 2021-10-22 | 华为技术有限公司 | Recommendation method and device based on automatic feature grouping |
| CN113536097B (en) * | 2020-04-14 | 2024-03-29 | 华为技术有限公司 | Recommendation method and device based on automatic feature grouping |
| CN111242752A (en) * | 2020-04-24 | 2020-06-05 | 支付宝(杭州)信息技术有限公司 | Method and system for determining recommended object based on multi-task prediction |
| CN113744009A (en) * | 2020-05-29 | 2021-12-03 | 北京沃东天骏信息技术有限公司 | Target object output method and device, computer readable medium and electronic equipment |
| CN111680221A (en) * | 2020-08-11 | 2020-09-18 | 腾讯科技（深圳）有限公司 | Information recommendation method, device, equipment and computer readable storage medium |
| CN112989202A (en) * | 2021-04-02 | 2021-06-18 | 常熟理工学院 | Personalized recommendation method and system based on dynamic network embedding |
| CN112989202B (en) * | 2021-04-02 | 2024-01-12 | 常熟理工学院 | Personalized recommendation method and system based on dynamic network embedding |
| CN113139856A (en) * | 2021-05-14 | 2021-07-20 | 青岛聚看云科技有限公司 | Movie and television member package recommendation method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110019163A (en) | Method, system, equipment and the storage medium of prediction, the recommendation of characteristics of objects | |
| AU2018383346B2 (en) | Domain-specific natural language understanding of customer intent in self-help | |
| US10402163B2 (en) | Intelligent data extraction | |
| US20170200205A1 (en) | Method and system for analyzing user reviews | |
| CN109948121A (en) | Article similarity method for digging, system, equipment and storage medium | |
| US20130073514A1 (en) | Flexible and scalable structured web data extraction | |
| CN107330613A (en) | A kind of public sentiment monitoring method, equipment and computer-readable recording medium | |
| EP3717984A1 (en) | Method and apparatus for providing personalized self-help experience | |
| CN107153656B (en) | Information searching method and device | |
| CN109472305A (en) | Answer quality determines model training method, answer quality determination method and device | |
| CN111198817A (en) | SaaS software fault diagnosis method and device based on convolutional neural network | |
| CN110019790A (en) | Text identification, text monitoring, data object identification, data processing method | |
| CN111625715B (en) | Information extraction method and device, electronic equipment and storage medium | |
| US10678821B2 (en) | Evaluating theses using tree structures | |
| CN109255012A (en) | A kind of machine reads the implementation method and device of understanding | |
| CN110458600A (en) | Portrait model training method, device, computer equipment and storage medium | |
| CN104077288B (en) | Web page contents recommend method and web page contents recommendation apparatus | |
| CN113837836A (en) | Model recommendation method, device, equipment and storage medium | |
| CN106575418A (en) | Suggested keywords | |
| CN109948040A (en) | Storage, recommended method and the system of object information, equipment and storage medium | |
| CN110262906B (en) | Interface label recommendation method and device, storage medium and electronic equipment | |
| CN107291686B (en) | Method and system for identifying emotion identification | |
| CN108460049A (en) | A kind of method and system of determining information category | |
| CN111611388A (en) | Account classification method, device and equipment | |
| CN107944589A (en) | The Forecasting Methodology and prediction meanss of ad click rate |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190716 |
|
| RJ01 | Rejection of invention patent application after publication |