CN110110139A - The method, apparatus and electronic equipment that a kind of pair of recommendation results explain - Google Patents
The method, apparatus and electronic equipment that a kind of pair of recommendation results explain Download PDFInfo
- Publication number
- CN110110139A CN110110139A CN201910317239.4A CN201910317239A CN110110139A CN 110110139 A CN110110139 A CN 110110139A CN 201910317239 A CN201910317239 A CN 201910317239A CN 110110139 A CN110110139 A CN 110110139A
- Authority
- CN
- China
- Prior art keywords
- disturbance
- sample
- feature
- probability
- prediction
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/70—Information retrieval; Database structures therefor; File system structures therefor of video data
- G06F16/73—Querying
- G06F16/735—Filtering based on additional data, e.g. user or group profiles
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/70—Information retrieval; Database structures therefor; File system structures therefor of video data
- G06F16/78—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Library & Information Science (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The method, apparatus and electronic equipment explained the embodiment of the invention provides a kind of pair of recommendation results, the described method includes: obtaining the corresponding prediction model of recommendation results to be explained and receiving the target user of the recommendation results to be explained, and initial sample is constructed, which includes: the content characteristic of recommendation included by recommendation results to be explained and the user characteristics of target user;Prediction model and initial sample are input in the preset local feature diagnostic model unrelated with prediction model, the importance values that each feature clicks prediction probability are obtained;According to the importance values, recommendation results to be explained are explained.It is a kind of model unrelated with prediction model, can be adapted for any prediction model due to preset local feature diagnostic model using the embodiment of the present invention.This improves the versatilities of means of interpretation.
Description
Technical field
The present invention relates to machine learning techniques field, the method explained more particularly to a kind of pair of recommendation results, dress
It sets and electronic equipment.
Background technique
Currently, many network application systems all can recommend some contents to user, to better meet user demand.Example
Such as: audio/video player system can recommend some videos for user, select for user.Specific recommended method mainly includes walking as follows
It is rapid:
Firstly, acquisition training sample, the personal information etc. of the video and user watched including user;Then,
Recommended models are trained with the sample of acquisition;Finally by the trained recommended models of the information input of candidate video and target user
In.Firstly, the prediction model in recommended models, from extracting feature in the information of candidate video and target user, comprising: video is special
It seeks peace user characteristics, candidate video is predicted by the probability that target user clicks, show that target user candidate regards to each
Probability is clicked in the prediction of frequency, and then, prediction is clicked the highest candidate video of probability as video to be recommended and pushed away by recommended models
It recommends to target user.
In order to be improved to prediction model, it is thus necessary to determine that each feature extracted is to prediction result, i.e., prediction is by point
The probability role hit, and then recommendation results are explained.
Currently, there are many common prediction models of recommender system, it can be general to clicking according to the feature of the sample of input
Rate is predicted.Such as: GBDT (Gradient Boosting Decision Tree, gradient promote decision tree) prediction model,
The prediction model or GBDT and LR that GBDT and FM (Factorization Machine, Factorization machine) are composed
The prediction model etc. that (Logistic Regression Classifier, logistic regression) is composed.
For tree-model as GBDT, prediction result can be analyzed using information gain method in the prior art,
Information gain method is to carry out judging characteristic by ranking results by being ranked up after each feature calculation information gain to sample
Importance, and then recommendation results are explained.Such as: feature A and feature B importance highest are obtained with information gain method, then
It is can be when explanation due to there is this 2 highest features of corresponding importance in video to be recommended, so the click probability of prediction is high.
However, inventor has found in the implementation of the present invention, at least there are the following problems for existing scheme:
Existing information gain method can only analyze the prediction result of tree-model as GBDT, cannot be to combination die
The prediction result of type is analyzed, such as: GBDT and the FM prediction model being composed or GBDT and LR cannot be composed
The prediction result of prediction model analyzed, and after the feature for introducing the above quantity of million ranks, information gain method
Also can not it is another one calculate characteristic information gain, can not determine the importance of each feature.
As it can be seen that the method explained in the prior art to recommendation results can not determine each some prediction models
The importance of feature, therefore all recommended models are not applied for, versatility is not high.
Summary of the invention
The embodiment of the present invention be designed to provide method, apparatus that a kind of pair of recommendation results explain and electronics is set
It is standby, to improve the versatility of means of interpretation.
In order to achieve the above objectives, the embodiment of the invention discloses the methods that a kind of pair of recommendation results explain, comprising:
Obtain the corresponding prediction model of recommendation results to be explained and target user;The recommendation results to be explained include to recommend
Content;The target user is the user for receiving the recommendation results to be explained;
Construct initial sample;The initial sample, comprising: the content characteristic of the recommendation and the user of target user
Feature;
By the prediction model and initial sample, it is input to the preset local feature diagnosis unrelated with the prediction model
In model;To in initial sample content characteristic and/or user characteristics repeatedly disturbed, obtain multiple disturbance samples;And it will
The initial sample is input to the prediction model, obtains corresponding initial predicted and clicks probability, each disturbance sample is distinguished
It is input to the prediction model, probability is clicked in prediction after obtaining corresponding multiple disturbances;And according to initial predicted click probability and
The difference between probability is clicked in prediction after each disturbance, obtains the importance values that each feature clicks prediction probability;Wherein,
Probability is clicked in the prediction are as follows: the target user that the prediction model is predicted according to input sample is to the recommendation
Click probability;
According to the importance values of each feature, the recommendation results to be explained are explained.
Optionally, the content characteristic and/or user characteristics in initial sample is repeatedly disturbed, and obtains multiple disturb
The step of dynamic sample, comprising:
From the content characteristic and user characteristics in initial sample, a feature being undisturbed is selected, as spy undetermined
Sign;
From the initial sample, the feature undetermined is hidden, primary disturbance is completed, obtains the corresponding disturbance of feature undetermined
Sample;Return described from the content characteristic and user characteristics in initial sample, select a feature being undisturbed, as to
The step of determining feature.
Optionally, described that the initial sample is input to the prediction model, it obtains corresponding initial predicted and clicks generally
The step of rate, comprising:
By the content characteristic and the user characteristics input prediction model in initial sample, corresponding initial predicted point is obtained
Hit probability;
It is described that each disturbance sample is separately input into the prediction model, it predicts to click after obtaining corresponding multiple disturbances
The step of probability, comprising:
The corresponding content characteristic disturbed in sample of each feature undetermined and user characteristics are inputted into the prediction model, obtained
Probability is clicked in prediction after obtaining the corresponding disturbance of each feature undetermined.
Optionally, described that difference between probability is clicked according to prediction after initial predicted click probability and each disturbance, it obtains
The step of importance values of probability are clicked for prediction to each feature, comprising:
Prediction after initial predicted clicks probability disturbance corresponding with each feature undetermined is calculated separately to click between probability
Difference;
Using each difference of acquisition as the importance values of corresponding feature undetermined.
Optionally, the prediction model includes: the first submodel and the second submodel;
By the prediction model and initial sample, it is input to the preset local feature unrelated with the prediction model and examines
Before step in disconnected model, further includes:
To the content characteristic and user characteristics in initial sample, by preset algorithm, all features are divided into dense characteristic
And sparse features;
The dense characteristic is input to the first submodel, obtains assemblage characteristic;
It is described that the initial sample is input to the prediction model, obtain the step that corresponding initial predicted clicks probability
Suddenly, comprising:
The assemblage characteristic and sparse features are inputted into second submodel, initial predicted is obtained and clicks probability.
Optionally, described by the prediction model and initial sample, it is input to preset unrelated with the prediction model
Step in local feature diagnostic model, comprising:
The assemblage characteristic and sparse features are spliced, spliced feature is obtained;
By the prediction model and the spliced feature, it is input to the preset part unrelated with the prediction model
In feature diagnostic model.
Optionally, the content characteristic and/or user characteristics in initial sample is repeatedly disturbed, and obtains multiple disturb
The step of dynamic sample, comprising:
The spliced feature is repeatedly disturbed, multiple disturbance samples are obtained;
It is described that each disturbance sample is separately input into the prediction model, it predicts to click after obtaining corresponding multiple disturbances
The step of probability, comprising:
Sample or the sparse spy of disturbance for each disturbance sample, from the disturbance sample, after obtaining disturbance dense characteristic
Sample after sign;
Sample after disturbance dense characteristic is input to the first submodel, obtains assemblage characteristic after the first disturbance;
By the sparse features input described the in assemblage characteristic after first disturbance and the sample after disturbance dense characteristic
Two submodels, probability is clicked in prediction after the corresponding disturbance of sample after obtaining the disturbance dense characteristic;
Or, the dense characteristic of the sample after disturbance sparse features is input to the first submodel, group after the second disturbance is obtained
Close feature;
By the sparse features input described the in assemblage characteristic after second disturbance and the sample after disturbance sparse features
Two submodels, probability is clicked in prediction after the corresponding disturbance of sample after obtaining the disturbance sparse features.
Optionally, first submodel are as follows: gradient promotes decision tree GBDT model, and second submodel returns for logic
Return LR model or Factorization machine FM model.
In order to achieve the above object, the embodiment of the invention also discloses the device that a kind of pair of recommendation results explain, packets
Include: first obtains module, building module, the second acquisition module and explanation module, wherein
Described first obtains module, for obtaining the corresponding prediction model of recommendation results to be explained and target user;It is described
Recommendation results to be explained include recommendation;The target user is the user for receiving the recommendation results to be explained;
The building module, for constructing initial sample;The initial sample, comprising: the content of the recommendation is special
It seeks peace the user characteristics of target user;
Described second obtains module, comprising: input submodule, disturbance submodule, probability obtain submodule, disturbance generally
Rate obtains submodule and importance values obtain submodule;
The input submodule, for being input to the preset and prediction mould for the prediction model and initial sample
In the unrelated local feature diagnostic model of type;
The disturbance submodule, for in initial sample content characteristic and/or user characteristics repeatedly disturbed, obtain
Obtain multiple disturbance samples;
The probability obtains submodule and is corresponded to for the initial sample to be input to the prediction model
Initial predicted click probability;
The disturbance probability obtains submodule, for each disturbance sample to be separately input into the prediction model, obtains
Probability is clicked in prediction after corresponding multiple disturbances;
The importance values obtain submodule, general for prediction click after clicking probability and each disturbance according to initial predicted
Difference between rate obtains the importance values that each feature clicks prediction probability;Wherein, probability is clicked in the prediction are as follows:
Click probability of the target user that the prediction model is predicted according to input sample to the recommendation;
The explanation module carries out the recommendation results to be explained for the importance values according to each feature
It explains.
Optionally, the disturbance submodule, comprising:
Selecting unit, for selecting a spy being undisturbed from the content characteristic and user characteristics in initial sample
Sign, as feature undetermined;
Unit is disturbed, for from the initial sample, hiding the feature undetermined, primary disturbance is completed, obtains undetermined
The corresponding disturbance sample of feature, triggers the selecting unit.
Optionally, the probability obtains submodule, specifically for by initial sample content characteristic and user it is special
Sign inputs the prediction model, obtains corresponding initial predicted and clicks probability;
The disturbance probability obtains submodule, specifically for the corresponding content disturbed in sample of each feature undetermined is special
User characteristics of seeking peace input the prediction model, and probability is clicked in prediction after obtaining the corresponding disturbance of each feature undetermined.
Optionally, the importance values obtain submodule, specifically for calculate separately initial predicted click probability with it is each
The difference between probability is clicked in prediction after the corresponding disturbance of feature undetermined;Using each difference of acquisition as corresponding feature undetermined
Importance values.
Optionally, the prediction model includes: the first submodel and the second submodel;
Described device, further includes:
Division module, for by the prediction model and initial sample, be input to it is preset with the prediction model without
Before in the local feature diagnostic model of pass, will own to the content characteristic and user characteristics in initial sample by preset algorithm
Feature is divided into dense characteristic and sparse features;
Input module obtains assemblage characteristic for the dense characteristic to be input to the first submodel;
The probability obtains submodule, is specifically used for the assemblage characteristic and sparse features input second son
Model obtains initial predicted and clicks probability.
Optionally, the input submodule, comprising:
Concatenation unit obtains spliced feature for splicing the assemblage characteristic and sparse features;
First input unit, for by the prediction model and the spliced feature, be input to it is preset with it is described
In the unrelated local feature diagnostic model of prediction model.
Optionally, the disturbance submodule obtains multiple specifically for repeatedly being disturbed to the spliced feature
Disturb sample;
The disturbance probability obtains submodule, comprising: obtaining unit, the second input unit and third input unit, or
It include: obtaining unit, the 4th input unit and the 5th input unit;
The obtaining unit, for from the disturbance sample, obtaining after disturbing dense characteristic for each disturbance sample
Sample after sample or disturbance sparse features;
Second input unit obtains first for the sample after disturbance dense characteristic to be input to the first submodel
Assemblage characteristic after disturbance;
The third input unit, for will it is described first disturbance after assemblage characteristic and disturbance dense characteristic after sample in
Sparse features input second submodel, prediction is clicked general after the corresponding disturbance of sample after obtaining the disturbance dense characteristic
Rate;
4th input unit, for the dense characteristic in the sample after disturbance sparse features to be input to the first submodule
Type obtains assemblage characteristic after the second disturbance;
5th input unit, for will it is described second disturbance after assemblage characteristic and disturbance sparse features after sample in
Sparse features input second submodel, prediction is clicked general after the corresponding disturbance of sample after obtaining the disturbance sparse features
Rate.
Optionally, first submodel are as follows: gradient promotes decision tree GBDT model, and second submodel returns for logic
Return LR model or Factorization machine FM model.
In order to achieve the above object, the electronics explained the embodiment of the invention also discloses a kind of pair of recommendation results is set
It is standby, including processor, communication interface, memory and communication bus, wherein the processor, the communication interface, the storage
Device completes mutual communication by the communication bus;
The memory, for storing computer program;
The processor when for executing the program stored on the memory, realizes any description above to recommendation
As a result the method and step explained.
At the another aspect that the present invention is implemented, a kind of computer readable storage medium is additionally provided, it is described computer-readable
Instruction is stored in storage medium, when run on a computer so that computer execute it is any of the above-described described to recommendation
As a result the method explained.
At the another aspect that the present invention is implemented, the embodiment of the invention also provides a kind of, and the computer program comprising instruction is produced
Product, when run on a computer, so that computer executes any of the above-described method explained to recommendation results.
As seen from the above technical solution, the method, apparatus provided in an embodiment of the present invention that recommendation results are explained and
Electronic equipment obtains the corresponding prediction model of recommendation results to be explained and receives the target user of the recommendation results to be explained, and
Initial sample is constructed, which includes: that the content characteristic of recommendation included by recommendation results to be explained and target are used
The user characteristics at family;Prediction model and initial sample are inputted into the preset local feature diagnostic model unrelated with prediction model
In, the importance values that each feature clicks prediction probability are obtained, according to the importance values, recommendation results to be explained are carried out
It explains.Due to preset local feature diagnostic model, it is a kind of model unrelated with prediction model, can be adapted for any prediction
Model.It is thus possible to improve the versatility of means of interpretation.
Certainly, it implements any of the products of the present invention or method must be not necessarily required to reach all the above excellent simultaneously
Point.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described.
Fig. 1 a is a kind of flow diagram of the method provided in an embodiment of the present invention explained to recommendation results;
Fig. 1 b is the schematic diagram of the treatment process of local feature diagnostic model in Fig. 1 a illustrated embodiment;
Fig. 2 is another flow diagram of the method provided in an embodiment of the present invention explained to recommendation results;
Fig. 3 is another flow diagram of the method provided in an embodiment of the present invention explained to recommendation results;
Fig. 4 is the structural schematic diagram of the device provided in an embodiment of the present invention explained to recommendation results;
Fig. 5 is the electronic equipment schematic diagram provided in an embodiment of the present invention explained to recommendation results.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention is described.
In order to solve prior art problem, the method that is explained the embodiment of the invention provides a kind of pair of recommendation results,
Device and electronic equipment.This method and device can be applied to various electronic equipments.The embodiment of the present invention is provided first below
The method that explains of a kind of pair of recommendation results be introduced.
As shown in Figure 1a, Fig. 1 a is a kind of process of the method provided in an embodiment of the present invention explained to recommendation results
Schematic diagram may include:
S101: the corresponding prediction model of recommendation results to be explained and target user are obtained;The recommendation results to be explained include
Recommendation;The target user is the user for receiving the recommendation results to be explained.
In practical applications, prediction model can be single model, such as GBDT or DNN (Deep Neural
Network, deep neural network) etc. models, be also possible to the built-up pattern of multiple models, such as the built-up pattern of GBDT and LR
Or built-up pattern of DNN and FM etc..Such as: for video recommendations, which be can be by inputting multiple use
The personal information of video and user that family had been watched and user click the probability for the video that these had been watched, and pass through
Repeatedly training, the prediction model that obtained training is completed.
S102: initial sample is constructed;The initial sample, comprising: the content characteristic of recommendation and the user of target user
Feature.
Such as: for video recommendations, which may include: feature and the target user of the video of recommendation
User characteristics.
S103: by prediction model and initial sample, it is input to the preset local feature diagnosis mould unrelated with prediction model
In type;To in initial sample content characteristic and/or user characteristics repeatedly disturbed, obtain multiple disturbance samples;And it will be first
Beginning sample is input to prediction model, obtains corresponding initial predicted and clicks probability, each disturbance sample is separately input into prediction
Model, probability is clicked in prediction after obtaining corresponding multiple disturbances;And it is predicted after according to initial predicted clicking probability and each disturbance
The difference between probability is clicked, the importance values that each feature clicks prediction probability are obtained.
In S103, the importance values that each feature clicks probability for prediction refer to: each feature pair of initial sample
The importance values of probability are clicked in prediction.
As shown in Figure 1 b, the treatment process of local feature diagnostic model includes:
1) initial sample is input in local feature diagnostic model, and local feature diagnostic model disturbs it, is obtained
Multiple disturbance samples;
Specifically, local feature diagnostic model in initial sample content characteristic and/or user characteristics repeatedly disturbed
It is dynamic, obtain multiple disturbance samples.
2) simultaneously, initial sample is also directly input to prediction model by local feature diagnostic model, is obtained corresponding initial
Probability is clicked in prediction.
3) then, each disturbance sample is separately input into prediction model by local feature diagnostic model, is obtained corresponding more
Probability is clicked in prediction after a disturbance.
4) probability is clicked in prediction after then, local diagnosis model clicks probability and each disturbance according to initial predicted, is calculated
Importance values of each feature of initial sample for prediction clicking rate.
Specifically, local feature diagnostic model, which can first calculate initial predicted, clicks future position after probability and each disturbance
The difference between probability is hit, predicts to click the difference between probability after then clicking probability and each disturbance according to initial predicted,
The each feature for obtaining initial sample clicks prediction the importance values of probability.
5) finally, local feature diagnostic model exports the importance values being calculated.
The above are the treatment processes of local feature diagnostic model, and process, can obtain the every of initial sample through this process
A feature clicks prediction the importance values of probability, so as to further the importance values be utilized to tie recommendation to be explained
Fruit explains.
S104: according to the importance values of each feature, recommendation results to be explained are explained.
The importance values of each feature of initial sample, it is big can to represent the contribution that this feature clicks probability for prediction
It is small.
Embodiment shown in application drawing 1a, the initial sample of the corresponding prediction model of recommendation results to be explained and building is defeated
Enter in the preset local feature diagnostic model unrelated with prediction model, obtains each feature and the important of probability is clicked for prediction
Property value explains recommendation results to be explained according to the importance values.It is one due to preset local feature diagnostic model
The kind model unrelated with prediction model, can be adapted for any prediction model.This improves the versatilities of means of interpretation.
Further, the embodiment of the invention provides another flow chart of the method explained to recommendation results,
In the embodiment, using local feature diagnostic model, select to be undisturbed from the content characteristic and user characteristics in initial sample
Feature, this feature is hidden, obtain disturbance sample, so that the importance values of each feature are obtained, to explain recommendation to be explained
As a result.Specifically, as shown in Fig. 2, may include:
S201: the corresponding prediction model of recommendation results to be explained and target user are obtained;The recommendation results to be explained include
Recommendation;The target user is the user for receiving the recommendation results to be explained.
Recommendation can be video, be also possible to audio or image etc., specifically without limitation.Recommendation knot to be explained
The corresponding prediction model of fruit can predict the probability that the recommendation results are clicked by target user.
S202: initial sample is constructed;The initial sample, comprising: the content characteristic of recommendation and the user of target user
Feature.
In one case, recommendation is video, then the content characteristic of recommendation, can be the label of video, such as
Performer, film, sitcom etc. or video length, video issuing date etc..The user characteristics of target user, can be use
The personal information at family, such as user's gender, age of user or user are in the number of registration etc. of audio/video player system.In this way, building
Initial sample can be the content characteristic and user characteristics for constructing initial sample.
In the present embodiment, one or more initial samples can be constructed.Specifically, when target user is one, it can be with
Using the user characteristics of the target user and the target user received recommendation results to be explained recommendation content it is special
Sign is to construct an initial sample.When target user is multiple, it can be directed to each target user, utilize the target user's
User characteristics and the target user the recommendation of received recommendation results to be explained content characteristic it is initial to construct one
Sample, to construct multiple initial samples.
S203: by prediction model and initial sample, it is input to the preset local feature diagnosis mould unrelated with prediction model
In type;To in initial sample content characteristic and/or user characteristics repeatedly disturbed, obtain multiple disturbance samples;It will be initial
Content characteristic and user characteristics input prediction model in sample obtain corresponding initial predicted and click probability;It will be each undetermined
Content characteristic and user characteristics input prediction model in the corresponding disturbance sample of feature, each feature undetermined of acquisition is corresponding to disturb
Probability is clicked in prediction after dynamic;Prediction after initial predicted clicks probability disturbance corresponding with each feature undetermined is calculated separately to click generally
Difference between rate;Using each difference of acquisition as the importance values of corresponding feature undetermined;Wherein, probability is clicked in prediction
Are as follows: click probability of the target user that prediction model is predicted according to input sample to recommendation.
In a kind of situation, if one initial sample of building, the feature of prediction model and initial sample can be formed
Vector inputs in above-mentioned local feature diagnostic model.
In another case, if the multiple initial samples of building, it can be defeated with a matrix type by multiple initial samples
Enter in above-mentioned local feature diagnostic model, in the matrix, one initial sample of each behavior, above-mentioned local feature diagnostic model
Data in matrix can be handled line by line.Here processing includes: the processing 1 in above-mentioned steps 103) -5).
In S203 in initial sample content characteristic and/or user characteristics repeatedly disturbed, obtain multiple disturbance samples
This, may include steps of:
1), from the content characteristic and user characteristics in initial sample, a feature being undisturbed is selected, as undetermined
Feature.
2), from initial sample, feature undetermined is hidden, primary disturbance is completed, obtains the corresponding disturbance sample of feature undetermined.
Wherein, feature undetermined is hidden, can be and set 0 for the value of feature undetermined.
3) it, returns from the content characteristic and user characteristics in initial sample, selects a feature being undisturbed, as
The step of feature undetermined.
By above three step, each feature in initial sample can once be disturbed, to obtain multiple
Disturb sample.
For example, it is assumed that the content characteristic and user characteristics in initial sample are five features of A, B, C, D, E, one is selected
A feature A being undisturbed hides feature A as feature undetermined, obtain the corresponding disturbance sample of feature A be [0, B, C, D,
E], then the reselection one feature B being undisturbed hides feature B as feature undetermined, obtains the corresponding disturbance of feature B
Sample is [A, 0, C, D, E], and so on, obtain multiple disturbance samples.
It, can be by the content characteristic and user characteristics input prediction mould in initial sample after obtaining multiple disturbance samples
Type obtains corresponding initial predicted and clicks probability;By the corresponding content characteristic disturbed in sample of each feature undetermined and user
Feature input prediction model, probability is clicked in prediction after obtaining the corresponding disturbance of each feature undetermined;Calculate separately initial predicted point
It hits prediction after probability disturbance corresponding with each feature undetermined and clicks difference between probability;Using each difference of acquisition as pair
The importance values for the feature undetermined answered.
Continue above-mentioned example, it is assumed that by the content characteristic and user characteristics A, B, C, D, E in initial sample, input prediction mould
Type, obtaining initial predicted and clicking probability is 0.6, and feature A is hidden, the corresponding disturbance sample of feature A [0, B, C, D, E] is inputted
Prediction model, prediction click probability is 0.1 after obtaining the corresponding disturbance of feature A, then calculates initial predicted and click probability and feature A
The difference between probability is clicked in prediction after corresponding disturbance, and obtaining the difference is 0.6-0.1=0.5, in this way, the difference 0.5 is just
It is the importance values of feature A undetermined.
Certainly, which is also possible to negative, for example, obtaining the corresponding disturbance sample of feature B after feature B undetermined is hidden
This [A, 0, C, D, E], by the disturbance sample input prediction model, prediction click probability is after obtaining the corresponding disturbance of feature B
0.8, then it calculates after initial predicted clicks probability disturbance corresponding with feature B and predicts to click the difference between probability, obtain the difference
Value is 0.6-0.8=-0.2, in this way, difference -0.2 is exactly the importance values of feature B undetermined.
Above-mentioned input sample refers to: the user characteristics of target user and the feature of Candidate Recommendation result.For example, Candidate Recommendation
As a result are as follows: candidate video.
S204: according to the importance values of each feature, recommendation results to be explained are explained.
Due to initial predicted click prediction after corresponding with the feature undetermined disturbance of probability click difference between probability can be with
It is positive number, negative or 0, and the difference just represents importance values, so, the importance values of each feature are also possible to positive number, bear
Number or 0.
The importance values can represent the contribution that feature clicks prediction probability.Importance values are bigger, corresponding
The contribution that feature clicks probability for prediction is bigger, on the contrary, importance values are smaller, corresponding feature clicks probability for prediction
Contribution it is smaller.
For example, it is assumed that the importance values of feature A are 0.1, the importance values of feature B are -0.3, then feature A treats solution
It releases recommendation results and plays positive contribution, and feature B plays negative sense contribution to recommendation results to be explained.
In a kind of situation, according to the importance values of each feature, recommendation results to be explained are explained, may is that pair
The importance values of each feature, are ranked up by size, according to ranking results, choose the maximum importance values pair of preset quantity
The feature answered explains recommendation results to be explained.
For example, the importance values of feature A are 0.1, the importance values of feature B are -0.3, and the importance values of feature C are 0.5,
The importance values of feature D are 0.8, then choose importance values maximum two features D and C, explain recommendation results to be explained.
In another case, can also be ranked up, the contribution margin of each feature according to contribution margin by contribution margin size
Size chooses the maximum feature of preset quantity contribution margin, explains recommendation results to be explained;Or it chooses and recommendation to be explained is tied
Fruit plays the feature of positive contribution, explains recommendation results to be explained.
For example, continuing above-mentioned example, the maximum feature D of contribution margin is chosen, explains recommendation results to be explained;Or selection pair
Recommendation results to be explained play feature D, C and A of positive contribution, explain recommendation results to be explained.
2 the embodiment described of application drawing is somebody's turn to do by obtaining the corresponding prediction model of recommendation results to be explained and receiving wait solve
The target user of recommendation results is released, and constructs initial sample, prediction model and initial sample are inputted into local feature diagnostic model
In;Intact feature is selected from the content characteristic and user characteristics in initial sample, this feature is hidden, obtains disturbance sample
This, to obtain initial sample and disturb the difference of sample, as the importance values for being hidden feature, to explain recommendation to be explained
As a result.Due to preset local feature diagnostic model, it is a kind of model unrelated with prediction model, can be adapted for any prediction
Model.This improves the versatilities of means of interpretation.
Further, the embodiment of the invention also provides another processes of the method explained to recommendation results
Scheme, in the embodiment, prediction model may include the first submodel and the second submodel, by all features in initial sample
It is combined and splices, obtain splicing feature, input in preset local feature diagnosis algorithm, obtain the importance of each feature
Value, to explain recommendation results.Specifically, as shown in figure 3, may include:
S301: the corresponding prediction model of recommendation results to be explained and target user are obtained;The recommendation results to be explained include
Recommendation;The target user is the user for receiving the recommendation results to be explained.
In a kind of situation, the corresponding prediction model of recommendation results to be explained, may include the first submodel and the second submodule
Type, the first submodel can be GBDT or DNN, for handling dense characteristic, not support processing higher-dimension sparse features;Second son
Model can be LR or FM, for handling sparse features.
S302: initial sample is constructed;The initial sample, comprising: the content characteristic of recommendation and the user of target user
Feature.
S303: all features are divided into dense by the content characteristic and user characteristics in initial sample by preset algorithm
Feature and sparse features.
For example, if initial sample includes the feature for the video recommended and the user characteristics of target user, this is first
May include in the feature of beginning sample the dense characteristics such as age of user, user's gender or video length and user number of registration or
The sparse features such as video tab.
After constructing initial sample, all features in initial sample can be numbered, by preset algorithm, will be owned
Feature is divided into dense characteristic and sparse features.Specifically, can be vector by each character representation in initial sample, it will be to
The corresponding feature of number for being less than preset threshold is divided into dense characteristic, will be greater than pre- by number of the length of amount as feature
If the corresponding feature of the number of threshold value is divided into sparse features.For example, preset threshold is 100,000, then the spy less than 100,000 will be numbered
Sign is determined as dense characteristic, and the feature by number greater than 100,000 is determined as sparse features.
S304: dense characteristic is input to the first submodel, obtains assemblage characteristic.
In practical applications, after dense characteristic being inputted the first submodel, dense characteristic can traverse the first submodel automatically
In all feature classifications all features in dense characteristic are automatically selected according to entropy production so that dense spy
All features of sign are divided into multiclass, and every one kind is an assemblage characteristic, and the feature discretization in every one kind is encoded, group is obtained
Feature vector is closed, so as to machine recognition processing.Since the feature quantity of initial sample is huge, the assemblage characteristic typically resulted in is height
Tie up assemblage characteristic.
It, can be by assemblage characteristic and sparse features input prediction model after obtaining assemblage characteristic in a kind of situation
Second submodel, the second submodel can calculate the feature of input, obtain initial predicted and click probability.
S305: assemblage characteristic and sparse features are spliced, and obtain spliced feature.
In a kind of situation, assemblage characteristic and sparse features are spliced, spliced feature is obtained, can be will combine
Feature generate feature vector, and by sparse features generate higher-dimension sparse features vector spliced, obtain splicing feature to
Amount.
For example, the feature vector of assemblage characteristic is [a1、a2、…an], higher-dimension sparse features vector is [b1、0、0、
0…b2], the feature vector of assemblage characteristic and higher-dimension sparse features vector are spliced, obtained splicing feature vector is [a1、
a2、…an、b1、0、0、0…b2]。
S306: it by prediction model and spliced feature, is input to the preset local feature unrelated with prediction model and examines
In disconnected model, the importance values that each feature clicks prediction probability are obtained.
In practical applications, using local feature diagnostic model, the spliced feature of input is repeatedly disturbed, is obtained
Obtain multiple disturbance samples;Sample or disturbance for each disturbance sample, from the disturbance sample, after obtaining disturbance dense characteristic
Sample after sparse features;
For example, continuing above-mentioned example, for the spliced feature [a of input1、a2、…an、b1、0、0、0…b2], it will be thick
Close feature A1Corresponding vector element a1It hides, obtains disturbance dense characteristic A1Sample [0, a afterwards2、…an、b1、0、0、0…
b2], or by sparse features B1Corresponding vector element b1It hides, obtains disturbance sparse features B1Sample [a afterwards1、a2、…an、
0、0、0、0…b2]。
Sample after disturbance dense characteristic is input to the first submodel, obtains assemblage characteristic after the first disturbance;By first
Assemblage characteristic and the sparse features in the sample after disturbance dense characteristic input the second submodel after disturbance, and it is dense to obtain the disturbance
Probability is clicked in prediction after the corresponding disturbance of sample after feature;
Specifically, the first submodel handles the dense characteristic in the sample after disturbance dense characteristic, first is obtained
Assemblage characteristic after disturbance, the second submodel is to the sparse spy in assemblage characteristic after the first disturbance and the sample after disturbance dense characteristic
Sign is handled, and probability is clicked in prediction after the corresponding disturbance of sample after obtaining the disturbance dense characteristic.
Alternatively, the dense characteristic in the sample after disturbance sparse features is input to the first submodel, the second disturbance is obtained
Assemblage characteristic afterwards;Sparse features in assemblage characteristic after second disturbance and the sample after disturbance sparse features are inputted into the second submodule
Type, probability is clicked in prediction after the corresponding disturbance of sample after obtaining the disturbance sparse features.
Specifically, the first submodel handles the dense characteristic in the sample after disturbance sparse features, second is obtained
Assemblage characteristic after disturbance, the second submodel is to the sparse spy in assemblage characteristic after the second disturbance and the sample after disturbance sparse features
Sign is handled, and probability is clicked in prediction after the corresponding disturbance of sample after obtaining the disturbance sparse features.
After probability is clicked in prediction after obtaining the corresponding disturbance of sample after disturbing dense characteristic, according to initial predicted point
The difference between probability is clicked in prediction after the corresponding disturbance of sample after hitting probability and each disturbance dense characteristic, alternatively, obtaining
After probability is clicked in prediction after the corresponding disturbance of sample after sparse features must be disturbed, probability and each is clicked according to initial predicted
The difference between probability is clicked in prediction after the corresponding disturbance of sample after disturbing sparse features, obtains each feature for future position
Hit the importance values of probability.Here, the importance values that each feature clicks probability for prediction refer to: in splicing feature vector
Each feature clicks prediction the importance values of probability.
S307: according to the importance values of each feature, recommendation results to be explained are explained.
It using embodiment shown in Fig. 3, is combined, spliced by the feature to initial sample, obtain splicing feature, it will
Splicing feature inputs in the preset local feature diagnostic model unrelated with prediction model, obtains each feature and prediction is clicked
The importance values of probability, and according to the importance values, explain recommendation results.It is one due to preset local feature diagnostic model
The kind model unrelated with prediction model, can be adapted for any prediction model.This improves the versatilities of means of interpretation.
Corresponding to embodiment of the method shown in Fig. 1 a, the embodiment of the invention also provides a kind of pair of recommendation results to explain
Device, as shown in figure 4, the apparatus may include: first obtain module 401, building module 402, second obtain module 403
With explanation module 404, wherein
Described first obtains module 401, for obtaining the corresponding prediction model of recommendation results to be explained and target user;Institute
Stating recommendation results to be explained includes recommendation;The target user is the user for receiving the recommendation results to be explained;
The building module 402, for constructing initial sample;The initial sample, comprising: the content of the recommendation
The user characteristics of feature and target user;
Described second obtains module 403, comprising: input submodule, disturbance submodule, probability obtain submodule, disturb
Dynamic probability obtains submodule and importance values obtain submodule;
The input submodule, for being input to the preset and prediction mould for the prediction model and initial sample
In the unrelated local feature diagnostic model of type;
The disturbance submodule, for in initial sample content characteristic and/or user characteristics repeatedly disturbed, obtain
Obtain multiple disturbance samples;
The probability obtains submodule and is corresponded to for the initial sample to be input to the prediction model
Initial predicted click probability;
The disturbance probability obtains submodule, for each disturbance sample to be separately input into the prediction model, obtains
Probability is clicked in prediction after corresponding multiple disturbances;
The importance values obtain submodule, general for prediction click after clicking probability and each disturbance according to initial predicted
Difference between rate obtains the importance values that each feature clicks prediction probability;Wherein, probability is clicked in the prediction are as follows:
Click probability of the target user that the prediction model is predicted according to input sample to the recommendation;
The explanation module 404, for the importance values according to each feature, to the recommendation results to be explained into
Row is explained.
Using embodiment shown in Fig. 4, obtains the corresponding prediction model of recommendation results to be explained and receive this and to be explained push away
The target user of result is recommended, and constructs initial sample, which includes: recommendation included by recommendation results to be explained
Content characteristic and target user user characteristics;Prediction model and initial sample are inputted preset unrelated with prediction model
In local feature diagnostic model, obtains the importance values that each feature clicks probability for prediction and treated according to the importance values
Explain that recommendation results explain.It is a kind of model unrelated with prediction model due to preset local feature diagnostic model, it can
To be suitable for any prediction model.This improves the versatilities of means of interpretation.
Specifically, in the present embodiment, the disturbance submodule may include:
Selecting unit, for selecting a spy being undisturbed from the content characteristic and user characteristics in initial sample
Sign, as feature undetermined;
Unit is disturbed, for from the initial sample, hiding the feature undetermined, primary disturbance is completed, obtains undetermined
The corresponding disturbance sample of feature, triggers the selecting unit.
Specifically, in the present embodiment, the probability obtains submodule, specifically for by the content in initial sample
Feature and user characteristics input the prediction model, obtain corresponding initial predicted and click probability;
The disturbance probability obtains submodule, specifically for the corresponding content disturbed in sample of each feature undetermined is special
User characteristics of seeking peace input the prediction model, and probability is clicked in prediction after obtaining the corresponding disturbance of each feature undetermined.
Specifically, in the present embodiment, the importance values obtain submodule, specifically for calculating separately initial predicted point
It hits prediction after probability disturbance corresponding with each feature undetermined and clicks difference between probability;Using each difference of acquisition as pair
The importance values for the feature undetermined answered.
Specifically, in the present embodiment, the prediction model includes: the first submodel and the second submodel;
Described device, further includes:
Division module (not shown), for being input to preset and institute by the prediction model and initial sample
Before stating in the unrelated local feature diagnostic model of prediction model, to the content characteristic and user characteristics in initial sample, by pre-
All features are divided into dense characteristic and sparse features by imputation method;
Input module (not shown) obtains assemblage characteristic for the dense characteristic to be input to the first submodel;
The probability obtains submodule, is specifically used for the assemblage characteristic and sparse features input second son
Model obtains initial predicted and clicks probability.
Specifically, in the present embodiment, the input submodule, comprising:
Concatenation unit obtains spliced feature for splicing the assemblage characteristic and sparse features;
First input unit, for by the prediction model and the spliced feature, be input to it is preset with it is described
In the unrelated local feature diagnostic model of prediction model.
Specifically, in the present embodiment, the disturbance submodule is specifically used for carrying out the spliced feature multiple
Disturbance, obtains multiple disturbance samples;
The disturbance probability obtains submodule, comprising: obtaining unit, the second input unit and third input unit, or
It include: obtaining unit, the 4th input unit and the 5th input unit;
The obtaining unit is used for for each disturbance sample, from the disturbance sample, after being disturbed after dense characteristic
Sample or disturbance after sample after sparse features;
Second input unit obtains first for the sample after disturbance dense characteristic to be input to the first submodel
Assemblage characteristic after disturbance;
The third input unit, for will it is described first disturbance after assemblage characteristic and disturbance dense characteristic after sample in
Sparse features input second submodel, prediction is clicked general after the corresponding disturbance of sample after obtaining the disturbance dense characteristic
Rate;
4th input unit, for the dense characteristic in the sample after disturbance sparse features to be input to the first submodule
Type obtains assemblage characteristic after the second disturbance;
5th input unit, for will it is described second disturbance after assemblage characteristic and disturbance sparse features after sample in
Sparse features input second submodel, prediction is clicked general after the corresponding disturbance of sample after obtaining the disturbance sparse features
Rate.
Specifically, in the present embodiment, first submodel are as follows: gradient promotion decision tree GBDT model, described second
Submodel is logistic regression LR model or Factorization machine FM model.
The embodiment of the invention also provides the electronic equipments that a kind of pair of recommendation results explain, as shown in figure 5, including place
Manage device 501, communication interface 502, memory 503 and communication bus 504, wherein processor 501, communication interface 502, memory
503 complete mutual communication by communication bus 504,
Memory 503, for storing computer program;
Processor 501 when for executing the program stored on memory 503, realizes following steps:
Obtain the corresponding prediction model of recommendation results to be explained and target user;The recommendation results to be explained include to recommend
Content;The target user is the user for receiving the recommendation results to be explained;
Construct initial sample;The initial sample, comprising: the content characteristic of the recommendation and the user of target user
Feature;
By the prediction model and initial sample, it is input to the preset local feature diagnosis unrelated with the prediction model
In model;To in initial sample content characteristic and/or user characteristics repeatedly disturbed, obtain multiple disturbance samples;And it will
The initial sample is input to the prediction model, obtains corresponding initial predicted and clicks probability, each disturbance sample is distinguished
It is input to the prediction model, probability is clicked in prediction after obtaining corresponding multiple disturbances;And according to initial predicted click probability and
The difference between probability is clicked in prediction after each disturbance, obtains the importance values that each feature clicks prediction probability;Wherein,
Probability is clicked in the prediction are as follows: the target user that the prediction model is predicted according to input sample is to the recommendation
Click probability;
According to the importance values of each feature, the recommendation results to be explained are explained.
As it can be seen that obtaining the corresponding prediction model of recommendation results to be explained in scheme provided by the embodiment of the present invention and connecing
The target user of the recommendation results to be explained is received, and constructs initial sample, which includes: that recommendation results to be explained are wrapped
The content characteristic of the recommendation included and the user characteristics of target user;The input of the feature of prediction model and initial sample is default
The local feature diagnostic model unrelated with prediction model in, obtain each feature for prediction click probability importance values,
According to the importance values, recommendation results to be explained are explained.It is a kind of and pre- due to preset local feature diagnostic model
The unrelated model of model is surveyed, can be adapted for any prediction model.This improves the versatilities of means of interpretation.
The communication bus that above-mentioned electronic equipment is mentioned can be Peripheral Component Interconnect standard (Peripheral Component
Interconnect, abbreviation PCI) bus or expanding the industrial standard structure (Extended Industry Standard
Architecture, abbreviation EISA) bus etc..The communication bus can be divided into address bus, data/address bus, control bus etc..
Only to be indicated with a thick line in figure, it is not intended that an only bus or a type of bus convenient for indicating.
Communication interface is for the communication between above-mentioned electronic equipment and other equipment.
Memory may include random access memory (Random Access Memory, abbreviation RAM), also may include
Nonvolatile memory (non-volatile memory), for example, at least a magnetic disk storage.Optionally, memory may be used also
To be storage device that at least one is located remotely from aforementioned processor.
Above-mentioned processor can be general processor, including central processing unit (Central Processing Unit,
Abbreviation CPU), network processing unit (Network Processor, abbreviation NP) etc.;It can also be digital signal processor
(Digital Signal Processing, abbreviation DSP), specific integrated circuit (Application Specific
Integrated Circuit, abbreviation ASIC), field programmable gate array (Field-Programmable Gate Array,
Abbreviation FPGA) either other programmable logic device, discrete gate or transistor logic, discrete hardware components.
In another embodiment provided by the invention, a kind of computer readable storage medium is additionally provided, which can
It reads to be stored with instruction in storage medium, when run on a computer, so that computer executes any institute in above-described embodiment
The method that recommendation results are explained stated, to obtain identical technical effect.
In another embodiment provided by the invention, a kind of computer program product comprising instruction is additionally provided, when it
When running on computers, so that computer executes any side explained to recommendation results in above-described embodiment
Method, to obtain identical technical effect.
In the above-described embodiments, can come wholly or partly by software, hardware, firmware or any combination thereof real
It is existing.When implemented in software, it can entirely or partly realize in the form of a computer program product.The computer program
Product includes one or more computer instructions.When loading on computers and executing the computer program instructions, all or
It partly generates according to process or function described in the embodiment of the present invention.The computer can be general purpose computer, dedicated meter
Calculation machine, computer network or other programmable devices.The computer instruction can store in computer readable storage medium
In, or from a computer readable storage medium to the transmission of another computer readable storage medium, for example, the computer
Instruction can pass through wired (such as coaxial cable, optical fiber, number from a web-site, computer, server or data center
User's line (DSL)) or wireless (such as infrared, wireless, microwave etc.) mode to another web-site, computer, server or
Data center is transmitted.The computer readable storage medium can be any usable medium that computer can access or
It is comprising data storage devices such as one or more usable mediums integrated server, data centers.The usable medium can be with
It is magnetic medium, (for example, floppy disk, hard disk, tape), optical medium (for example, DVD) or semiconductor medium (such as solid state hard disk
Solid State Disk (SSD)) etc..
It should be noted that, in this document, relational terms such as first and second and the like are used merely to a reality
Body or operation are distinguished with another entity or operation, are deposited without necessarily requiring or implying between these entities or operation
In any actual relationship or order or sequence.Moreover, the terms "include", "comprise" or its any other variant are intended to
Non-exclusive inclusion, so that the process, method, article or equipment including a series of elements is not only wanted including those
Element, but also including other elements that are not explicitly listed, or further include for this process, method, article or equipment
Intrinsic element.In the absence of more restrictions, the element limited by sentence "including a ...", it is not excluded that
There is also other identical elements in process, method, article or equipment including the element.
Each embodiment in this specification is all made of relevant mode and describes, same and similar portion between each embodiment
Dividing may refer to each other, and each embodiment focuses on the differences from other embodiments.Especially for device reality
For applying example, apparatus embodiments and storage medium embodiment, since it is substantially similar to the method embodiment, so the comparison of description
Simply, the relevent part can refer to the partial explaination of embodiments of method.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the scope of the present invention.It is all
Any modification, equivalent replacement, improvement and so within the spirit and principles in the present invention, are all contained in protection scope of the present invention
It is interior.
Claims (17)
1. the method that a kind of pair of recommendation results explain, which is characterized in that the described method includes:
Obtain the corresponding prediction model of recommendation results to be explained and target user;The recommendation results to be explained include in recommending
Hold;The target user is the user for receiving the recommendation results to be explained;
Construct initial sample;The initial sample, comprising: the content characteristic of the recommendation and the user of target user are special
Sign;
By the prediction model and initial sample, it is input to the preset local feature diagnostic model unrelated with the prediction model
In;To in initial sample content characteristic and/or user characteristics repeatedly disturbed, obtain multiple disturbance samples;And it will be described
Initial sample is input to the prediction model, obtains corresponding initial predicted and clicks probability, each disturbance sample is inputted respectively
To the prediction model, probability is clicked in prediction after obtaining corresponding multiple disturbances;And probability and each is clicked according to initial predicted
The difference between probability is clicked in prediction after disturbance, obtains the importance values that each feature clicks prediction probability;Wherein, described
Probability is clicked in prediction are as follows: click of the target user that the prediction model is predicted according to input sample to the recommendation
Probability;
According to the importance values of each feature, the recommendation results to be explained are explained.
2. the method according to claim 1, wherein
The content characteristic and/or user characteristics in initial sample is repeatedly disturbed, and the step of multiple disturbance samples is obtained
Suddenly, comprising:
From the content characteristic and user characteristics in initial sample, a feature being undisturbed is selected, as feature undetermined;
From the initial sample, the feature undetermined is hidden, primary disturbance is completed, obtains the corresponding disturbance sample of feature undetermined
This;Return to described from the content characteristic and user characteristics in initial sample, one feature being undisturbed of selection, as undetermined
The step of feature.
3. according to the method described in claim 2, it is characterized in that,
It is described that the initial sample is input to the prediction model, obtain the step of corresponding initial predicted clicks probability, packet
It includes:
By the content characteristic and the user characteristics input prediction model in initial sample, obtains corresponding initial predicted and click generally
Rate;
Described that each disturbance sample is separately input into the prediction model, probability is clicked in prediction after obtaining corresponding multiple disturbances
The step of, comprising:
The corresponding content characteristic disturbed in sample of each feature undetermined and user characteristics are inputted into the prediction model, obtained each
Probability is clicked in prediction after the corresponding disturbance of a feature undetermined.
4. according to the method described in claim 3, it is characterized in that,
It is described that difference between probability is clicked according to prediction after initial predicted click probability and each disturbance, obtain each feature pair
In the step of importance values of probability are clicked in prediction, comprising:
It calculates separately initial predicted and clicks prediction after probability disturbance corresponding with each feature undetermined and click difference between probability;
Using each difference of acquisition as the importance values of corresponding feature undetermined.
5. the method according to claim 1, wherein the prediction model, includes: the first submodel and the second son
Model;
By the prediction model and initial sample, it is input to the preset local feature diagnosis mould unrelated with the prediction model
Before step in type, further includes:
To the content characteristic and user characteristics in initial sample, by preset algorithm, all features are divided into dense characteristic and dilute
Dredge feature;
The dense characteristic is input to the first submodel, obtains assemblage characteristic;
It is described that the initial sample is input to the prediction model, obtain the step of corresponding initial predicted clicks probability, packet
It includes:
The assemblage characteristic and sparse features are inputted into second submodel, initial predicted is obtained and clicks probability.
6. according to the method described in claim 5, it is characterized in that,
It is described by the prediction model and initial sample, be input to the preset local feature diagnosis unrelated with the prediction model
Step in model, comprising:
The assemblage characteristic and sparse features are spliced, spliced feature is obtained;
By the prediction model and the spliced feature, it is input to the preset local feature unrelated with the prediction model
In diagnostic model.
7. according to the method described in claim 6, it is characterized in that,
The content characteristic and/or user characteristics in initial sample is repeatedly disturbed, and the step of multiple disturbance samples is obtained
Suddenly, comprising:
The spliced feature is repeatedly disturbed, multiple disturbance samples are obtained;
Described that each disturbance sample is separately input into the prediction model, probability is clicked in prediction after obtaining corresponding multiple disturbances
The step of, comprising:
For each disturbance sample, from the disturbance sample, after the sample or disturbance sparse features after obtaining disturbance dense characteristic
Sample;
Sample after disturbance dense characteristic is input to the first submodel, obtains assemblage characteristic after the first disturbance;
By sparse features input second son in assemblage characteristic after first disturbance and the sample after disturbance dense characteristic
Model, probability is clicked in prediction after the corresponding disturbance of sample after obtaining the disturbance dense characteristic;
Or, the dense characteristic of the sample after disturbance sparse features is input to the first submodel, spy is combined after obtaining the second disturbance
Sign;
By sparse features input second son in assemblage characteristic after second disturbance and the sample after disturbance sparse features
Model, probability is clicked in prediction after the corresponding disturbance of sample after obtaining the disturbance sparse features.
8. according to the described in any item methods of claim 5~7, which is characterized in that first submodel are as follows: gradient promotion is determined
Plan tree GBDT model, second submodel are logistic regression LR model or Factorization machine FM model.
9. the device that a kind of pair of recommendation results explain, which is characterized in that described device includes: the first acquisition module, building
Module, second obtain module and explanation module, wherein
Described first obtains module, for obtaining the corresponding prediction model of recommendation results to be explained and target user;It is described wait solve
Releasing recommendation results includes recommendation;The target user is the user for receiving the recommendation results to be explained;
The building module, for constructing initial sample;The initial sample, comprising: the content characteristic of the recommendation and
The user characteristics of target user;
Described second obtains module, comprising: input submodule, disturbance submodule, probability obtain submodule, disturbance probability obtains
It obtains submodule and importance values obtains submodule;
The input submodule, for by the prediction model and initial sample, be input to it is preset with the prediction model without
In the local feature diagnostic model of pass;
The disturbance submodule, for in initial sample content characteristic and/or user characteristics repeatedly disturbed, obtain it is more
A disturbance sample;
The probability obtains submodule, for the initial sample to be input to the prediction model, obtains corresponding first
Begin prediction click probability;
The disturbance probability obtains submodule and is corresponded to for each disturbance sample to be separately input into the prediction model
Multiple disturbances after prediction click probability;
The importance values obtain submodule, for according to initial predicted click prediction after probability and each disturbance click probability it
Between difference, obtain each feature for prediction click probability importance values;Wherein, probability is clicked in the prediction are as follows: described
Click probability of the target user that prediction model is predicted according to input sample to the recommendation;
The explanation module explains the recommendation results to be explained for the importance values according to each feature.
10. device according to claim 9, which is characterized in that the disturbance submodule, comprising:
Selecting unit is made for selecting a feature being undisturbed from the content characteristic and user characteristics in initial sample
For feature undetermined;
Unit is disturbed, for from the initial sample, hiding the feature undetermined, primary disturbance is completed, obtains feature undetermined
Corresponding disturbance sample, triggers the selecting unit.
11. device according to claim 10, which is characterized in that
The probability obtains submodule, specifically for by initial sample content characteristic and user characteristics input it is described pre-
Model is surveyed, corresponding initial predicted is obtained and clicks probability;
The disturbance probability obtains submodule, specifically for by the corresponding disturbance sample of each feature undetermined content characteristic and
User characteristics input the prediction model, and probability is clicked in prediction after obtaining the corresponding disturbance of each feature undetermined.
12. device according to claim 11, which is characterized in that
The importance values obtain submodule, corresponding with each feature undetermined specifically for calculating separately initial predicted click probability
Disturbance after prediction click probability between difference;Using each difference of acquisition as the importance values of corresponding feature undetermined.
13. device according to claim 9, which is characterized in that the prediction model includes: the first submodel and second
Submodel;
Described device, further includes:
Division module, for being input to preset unrelated with the prediction model by the prediction model and initial sample
Before in local feature diagnostic model, to the content characteristic and user characteristics in initial sample, by preset algorithm, by all features
It is divided into dense characteristic and sparse features;
Input module obtains assemblage characteristic for the dense characteristic to be input to the first submodel;
The probability obtains submodule, is specifically used for the assemblage characteristic and sparse features inputting second submodule
Type obtains initial predicted and clicks probability.
14. device according to claim 13, which is characterized in that the input submodule, comprising:
Concatenation unit obtains spliced feature for splicing the assemblage characteristic and sparse features;
First input unit, for being input to the preset and prediction for the prediction model and the spliced feature
In the unrelated local feature diagnostic model of model.
15. device according to claim 14, which is characterized in that
The disturbance submodule obtains multiple disturbance samples specifically for repeatedly being disturbed to the spliced feature;
The disturbance probability obtains submodule, comprising: obtaining unit, the second input unit and third input unit, or include:
Obtaining unit, the 4th input unit and the 5th input unit;
The obtaining unit, for the sample for each disturbance sample, from the disturbance sample, after obtaining disturbance dense characteristic
Or the sample after disturbance sparse features;
Second input unit obtains the first disturbance for the sample after disturbance dense characteristic to be input to the first submodel
Assemblage characteristic afterwards;
The third input unit, for will it is described first disturbance after assemblage characteristic and disturbance dense characteristic after sample in it is dilute
It dredges feature and inputs second submodel, probability is clicked in prediction after the corresponding disturbance of sample after obtaining the disturbance dense characteristic;
4th input unit, for the dense characteristic in the sample after disturbance sparse features to be input to the first submodel,
Assemblage characteristic after the second disturbance of acquisition;
5th input unit, for will it is described second disturbance after assemblage characteristic and disturbance sparse features after sample in it is dilute
It dredges feature and inputs second submodel, probability is clicked in prediction after the corresponding disturbance of sample after obtaining the disturbance sparse features.
16. 3~15 described in any item devices according to claim 1, which is characterized in that first submodel are as follows: gradient mentions
Decision tree GBDT model is risen, second submodel is logistic regression LR model or Factorization machine FM model.
17. a kind of electronic equipment, which is characterized in that including processor, communication interface, memory and communication bus, wherein described
Processor, the communication interface, the memory complete mutual communication by the communication bus;
The memory, for storing computer program;
The processor when for executing the program stored on the memory, realizes any side claim 1-8
Method step.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910317239.4A CN110110139B (en) | 2019-04-19 | 2019-04-19 | Method and device for explaining recommendation result and electronic equipment |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910317239.4A CN110110139B (en) | 2019-04-19 | 2019-04-19 | Method and device for explaining recommendation result and electronic equipment |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110110139A true CN110110139A (en) | 2019-08-09 |
CN110110139B CN110110139B (en) | 2021-06-22 |
Family
ID=67485893
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910317239.4A Active CN110110139B (en) | 2019-04-19 | 2019-04-19 | Method and device for explaining recommendation result and electronic equipment |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110110139B (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110660485A (en) * | 2019-08-20 | 2020-01-07 | 南京医渡云医学技术有限公司 | Method and device for acquiring influence of clinical index |
CN111062442A (en) * | 2019-12-20 | 2020-04-24 | 支付宝(杭州)信息技术有限公司 | Method and device for explaining service processing result of service processing model |
CN111861253A (en) * | 2020-07-29 | 2020-10-30 | 北京车薄荷科技有限公司 | Personnel capacity determining method and system |
CN112308132A (en) * | 2020-10-29 | 2021-02-02 | 中山大学 | Robust training method and system based on factorization machine |
CN112750529A (en) * | 2020-12-31 | 2021-05-04 | 平安科技(深圳)有限公司 | Intelligent medical inquiry device, equipment and medium |
CN113486242A (en) * | 2021-07-13 | 2021-10-08 | 同济大学 | Non-invasive personalized interpretation method and system based on recommendation system |
CN114841233A (en) * | 2022-03-22 | 2022-08-02 | 阿里巴巴(中国)有限公司 | Path interpretation method, device and computer program product |
CN116645211A (en) * | 2023-05-15 | 2023-08-25 | 中信建投证券股份有限公司 | Recommended user information generation method, apparatus, device and computer readable medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105095477A (en) * | 2015-08-12 | 2015-11-25 | 华南理工大学 | Recommendation algorithm based on multi-index grading |
US20170193589A1 (en) * | 2009-04-27 | 2017-07-06 | Hewlett Packard Enterprise Development Lp | System And Method For Making A Recommendation Based On User Data |
CN108198045A (en) * | 2018-01-30 | 2018-06-22 | 东华大学 | The design method of mixing commending system based on e-commerce website data mining |
CN108460619A (en) * | 2018-01-16 | 2018-08-28 | 杭州电子科技大学 | A kind of fusion shows the Collaborative Recommendation model of implicit feedback |
-
2019
- 2019-04-19 CN CN201910317239.4A patent/CN110110139B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170193589A1 (en) * | 2009-04-27 | 2017-07-06 | Hewlett Packard Enterprise Development Lp | System And Method For Making A Recommendation Based On User Data |
CN105095477A (en) * | 2015-08-12 | 2015-11-25 | 华南理工大学 | Recommendation algorithm based on multi-index grading |
CN108460619A (en) * | 2018-01-16 | 2018-08-28 | 杭州电子科技大学 | A kind of fusion shows the Collaborative Recommendation model of implicit feedback |
CN108198045A (en) * | 2018-01-30 | 2018-06-22 | 东华大学 | The design method of mixing commending system based on e-commerce website data mining |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110660485A (en) * | 2019-08-20 | 2020-01-07 | 南京医渡云医学技术有限公司 | Method and device for acquiring influence of clinical index |
CN111062442A (en) * | 2019-12-20 | 2020-04-24 | 支付宝(杭州)信息技术有限公司 | Method and device for explaining service processing result of service processing model |
CN111062442B (en) * | 2019-12-20 | 2022-04-12 | 支付宝(杭州)信息技术有限公司 | Method and device for explaining service processing result of service processing model |
CN111861253A (en) * | 2020-07-29 | 2020-10-30 | 北京车薄荷科技有限公司 | Personnel capacity determining method and system |
CN112308132A (en) * | 2020-10-29 | 2021-02-02 | 中山大学 | Robust training method and system based on factorization machine |
CN112750529A (en) * | 2020-12-31 | 2021-05-04 | 平安科技(深圳)有限公司 | Intelligent medical inquiry device, equipment and medium |
CN113486242A (en) * | 2021-07-13 | 2021-10-08 | 同济大学 | Non-invasive personalized interpretation method and system based on recommendation system |
CN114841233A (en) * | 2022-03-22 | 2022-08-02 | 阿里巴巴(中国)有限公司 | Path interpretation method, device and computer program product |
CN114841233B (en) * | 2022-03-22 | 2024-05-31 | 阿里巴巴(中国)有限公司 | Path interpretation method, apparatus and computer program product |
CN116645211A (en) * | 2023-05-15 | 2023-08-25 | 中信建投证券股份有限公司 | Recommended user information generation method, apparatus, device and computer readable medium |
CN116645211B (en) * | 2023-05-15 | 2024-05-10 | 中信建投证券股份有限公司 | Recommended user information generation method, apparatus, device and computer readable medium |
Also Published As
Publication number | Publication date |
---|---|
CN110110139B (en) | 2021-06-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110110139A (en) | The method, apparatus and electronic equipment that a kind of pair of recommendation results explain | |
US10643109B2 (en) | Method and system for automatically classifying data expressed by a plurality of factors with values of text word and symbol sequence by using deep learning | |
Gulli et al. | Deep learning with Keras | |
US20200380023A1 (en) | Classifying data objects | |
CN110209922B (en) | Object recommendation method and device, storage medium and computer equipment | |
CN111741330B (en) | Video content evaluation method and device, storage medium and computer equipment | |
CN109325146B (en) | Video recommendation method and device, storage medium and server | |
CN111046286A (en) | Object recommendation method and device and computer storage medium | |
CN109101481A (en) | A kind of name entity recognition method, device and electronic equipment | |
CN110234018B (en) | Multimedia content description generation method, training method, device, equipment and medium | |
CN105144164A (en) | Scoring concept terms using a deep network | |
US20220245424A1 (en) | Microgenre-based hyper-personalization with multi-modal machine learning | |
CN112732870B (en) | Word vector based search method, device, equipment and storage medium | |
CN111010592B (en) | Video recommendation method and device, electronic equipment and storage medium | |
US10353951B1 (en) | Search query refinement based on user image selections | |
CN110717099A (en) | Method and terminal for recommending film | |
CN110209875B (en) | User content portrait determination method, access object recommendation method and related device | |
CN109657044A (en) | Data retrieval method, data reordering method, device, terminal and storage medium | |
CN112989146B (en) | Method, apparatus, device, medium and program product for recommending resources to target user | |
CN112182281B (en) | Audio recommendation method, device and storage medium | |
CN113033707B (en) | Video classification method and device, readable medium and electronic equipment | |
CN112989118B (en) | Video recall method and device | |
CN114428910A (en) | Resource recommendation method and device, electronic equipment, product and medium | |
CN113869377A (en) | Training method and device and electronic equipment | |
KR20190108958A (en) | Method and Apparatus for Explicit Lyrics Classification Using Automated Explicit Lexicon Generation and Machine Learning |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |