WO2020233432A1

WO2020233432A1 - Method and device for information recommendation

Info

Publication number: WO2020233432A1
Application number: PCT/CN2020/089541
Authority: WO
Inventors: 刘家豪; 谢淼; 彭艺; 王寅; 王超; 李楠; 杨程
Original assignee: 阿里巴巴集团控股有限公司
Priority date: 2019-05-20
Filing date: 2020-05-11
Publication date: 2020-11-26
Also published as: CN111967892A

Abstract

A method and device for information recommendation. The method comprises: acquiring user behavior data and processing the acquired user behavior data to obtain user attribute information and user behavior feedback information (100); updating a parameter item and a non-parameter item of an evaluation model on the basis of the user behavior feedback information of a user with respect to a recommended candidate item, and generating, on the basis of the user attribute information and of the user behavior feedback information, user characteristic information expressing a user preference (101); when making an information recommendation to the user once again, evaluating, on the basis of the updated evaluation model and of the user characteristic information, each candidate item in a candidate dataset to be recommended, recommending a candidate item on the basis of the evaluation result, and returning to step 101. The method combines the advantages of a parameter model and of a non-parameter model, implements fast convergence, and ensures a shortened theoretical loss distance from an optimal solution, thus ensuring that a recommended result is of increased quality, and enhancing user experience.

Description

Information recommendation method and device

This application claims the priority of the Chinese patent application filed on May 20, 2019, with the application number 201910418493.3 and the invention title "A method and device for information recommendation", the entire content of which is incorporated into this application by reference.

Technical field

This application relates to but is not limited to e-commerce technology, especially an information recommendation method and device.

Background technique

With the rapid development of the Internet, especially the development of the mobile Internet, recommendation systems are playing an increasingly important role. In the recommendation system, the requirement of some scenarios is to select an optimal candidate information from the candidate data set and recommend it to the user. For example, in the field of e-commerce, one best cover image is recommended for a product from multiple cover images; another example is for service optimization issues, an optimal hotspot from multiple WiFi hotspots will be selected and recommended to users ; For another example, for the problem of selecting a video cover image, a best image for the same video from multiple images will be recommended as the cover image of the video, and so on.

For these application scenarios, the method of recommending candidates in related technologies is only suitable for parametric modeling scenarios, that is, modeling as a sorting problem. For each candidate information in the candidate data set, use the parameterized model to calculate separately according to the characteristics Based on an estimated value, an optimal candidate information is selected and recommended to the user based on the obtained estimated value.

Take the example of choosing a cover image for a product. In the field of e-commerce, businesses will release a lot of products and also design or shoot a lot of promotional images for the products. These propaganda images themselves are different in quality. In the related technology, according to the difference in the quality of the promotional image itself, the quality parameter is used to select the best promotional image for the product as the cover image of the product. However, there is a problem with this selection method. Different users have their own preferences and preferences for different pictures. In this way, the cover image selected only from the picture quality may not meet the needs of the user, that is, the related technology provides The inaccurate information recommendation program will directly cause the user to experience poor user experience when browsing the product, thereby reducing the transaction conversion rate of the product.

Summary of the invention

The present application provides an information recommendation method and device, which can implement high-quality information recommendation, thereby improving user experience.

The embodiment of the present invention provides an information recommendation method, including:

Update the parameter items and non-parameter items of the evaluation model based on the user behavior feedback information of the recommended candidates, and generate user characteristic information that represents the user's preferences based on the user attribute information and user behavior feedback information;

When information recommendation is made to the user again, each candidate in the candidate data set to be recommended is evaluated according to the updated evaluation model and user characteristic information, the candidate is recommended according to the evaluation result, and the update step is returned.

In an exemplary embodiment, the method further includes:

The candidate data set to be recommended is sorted to reduce the size of the candidate data set to be recommended.

In an exemplary embodiment, the method further includes:

Recommending the recommended candidate to the user from the candidate data set to be recommended.

In an exemplary embodiment, the updating the parameter items of the evaluation model according to user behavior feedback information includes:

Updating the parameter matrix according to the parameter matrix when the recommended candidate is recommended and the characteristic information of the recommended candidate;

A new parameter vector is determined according to the updated parameter matrix, the characteristic information of the recommended candidate and the user behavior feedback information.

In an exemplary embodiment, the updating the non-parameter items of the model according to user behavior feedback information includes:

Determine the new non-parametric estimated value of the recommended candidate according to the recommended times of the recommended candidate, the updated recommended times and the non-parametric estimated value of the recommended candidate;

Determine the new feature vector of the recommended candidate according to the number of recommendations of the recommended candidate, the updated number of recommendations, the mean value of the feature vector of the recommended candidate, and the feature information of the recommended candidate Mean.

In an exemplary embodiment, the generating user characteristic information that represents user preferences includes:

For each candidate item in the candidate data set to be recommended, feature information of the candidate item is generated according to the intersection of the user attribute information and the feature of the candidate item to construct the feature set representing user feature information that the user likes .

In an exemplary embodiment, the evaluating each candidate in the candidate data set to be recommended according to the updated evaluation model and user characteristic information, and recommending the candidates according to the evaluation result, includes:

The updated parameter items, the non-parameter items, and the generated user characteristic information are used as the input of the evaluation model, and the candidates in the candidate data set to be recommended are scored and sorted, and the one with the highest score The candidate items are recommended to the user as the candidates when the information is recommended to the user again.

In an exemplary embodiment, the scoring and sorting the candidates in the candidate data set to be recommended, and using the candidate with the highest score as the candidate for the second information recommendation to the user includes: :

For each candidate in the candidate data set to be recommended, the value of each candidate is calculated according to the model parameter vector information, the feature vector information of the candidate, the user behavior feedback information for the candidate, and the average user behavior feedback. Estimated value P _t (a);

The estimated value of each candidate set P _t (a), for each of the candidates are sorted, the maximum value of the estimated candidates as a _t the candidate again when information recommendation to the recommendation to the user述Users.

The present application also provides a computer-readable storage medium that stores computer-executable instructions, and the computer-executable instructions are used to execute the information recommendation method described in any one of the above.

The present application further provides a device for implementing named entity recognition, including a memory and a processor, wherein the memory stores the following instructions that can be executed by the processor: for executing the information recommendation method described in any one of the above A step of.

This application further provides an information recommendation device, including: a training module, a generating module, and an evaluation module; wherein,

The training module is used to update the parameter items and non-parameter items of the evaluation model according to the user behavior feedback information of the recommended candidates;

The generating module is used for each candidate in the candidate data set to be recommended, and generate user characteristic information that represents the user's preferences based on user attribute information and user behavior feedback information of the user on the recommended candidate;

The evaluation module is used to evaluate each candidate in the candidate data set to be recommended according to the updated evaluation model and user characteristic information when information is recommended to the user again, and recommend the candidates according to the evaluation result.

In an illustrative example, the device further includes:

The collection module is used to obtain user behavior data, and process the obtained user behavior data to obtain the user attribute information and the user behavior feedback information.

In an illustrative example, the device further includes:

The recall module is used to obtain the candidate data set to be recommended from the database.

In an exemplary embodiment, the recall module is also used to:

This application includes: updating the parameter items and non-parameter items of the evaluation model according to the user behavior feedback information of the recommended candidates, and generating user characteristic information indicating the user’s preferences based on the user attribute information and user behavior feedback information; When the user makes information recommendation, evaluate each candidate in the candidate data set to be recommended according to the updated evaluation model and user characteristic information, recommend the candidates according to the evaluation result, and return the parameter items and non-parameter items of the evaluation model Steps to update. This application introduces non-parametric items through the semi-parametric environment, so that the evaluation model can fit the distribution of unknown data, achieve high-quality information recommendation, and improve user experience.

Other features and advantages of the present invention will be described in the following description, and partly become obvious from the description, or understood by implementing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the structures specifically pointed out in the specification, claims and drawings.

Description of the drawings

The accompanying drawings are used to provide a further understanding of the technical solution of the present application, and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the technical solution of the present application, and do not constitute a limitation to the technical solution of the present application.

FIG. 1 is a schematic flowchart of an embodiment of an application information recommendation method;

FIG. 2 is a schematic diagram of the composition structure of an embodiment of the information recommendation device of this application.

Detailed ways

In order to make the purpose, technical solutions, and advantages of the present application clearer, the embodiments of the present application will be described in detail below in conjunction with the accompanying drawings. It should be noted that the embodiments in this application and the features in the embodiments can be combined with each other arbitrarily if there is no conflict.

In a typical configuration of this application, the computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.

The memory may include non-permanent memory in computer readable media, random access memory (RAM) and/or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM). Memory is an example of computer readable media.

Computer-readable media include permanent and non-permanent, removable and non-removable media, and information storage can be realized by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disc (DVD) or other optical storage, Magnetic cassettes, magnetic tape storage or other magnetic storage devices or any other non-transmission media can be used to store information that can be accessed by computing devices. According to the definition in this article, computer-readable media does not include non-transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

The steps shown in the flowchart of the drawings may be executed in a computer system such as a set of computer-executable instructions. And, although a logical sequence is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than here.

Taking the recommended scene of the cover image as an example (other scenes are similar), the inventor of the present application believes that in the process of selecting a picture, the factors that determine the quality of the picture include two aspects: one is the quality of the picture itself, which is an inherent attribute of the picture. The quality is a fixed value, which is determined when the picture is generated, and it will not change with time and environment. It can be called a non-parametric item; on the other hand, the user's preference will change with time and environment. It is called a parameter item. Therefore, for issues such as the recommendation of the best cover image of a product or video, it can be regarded as a problem of image selection in a semi-parametric environment, where semi-parametric refers to the combination of a parameter part and a non-parameter part.

The change factor reflecting the user’s preferences can be regarded as a parameter item. In the embodiment of this application, it is proposed that a feature vector can be used to express it, and the user’s preference can be obtained through model learning, as shown in the evaluation of formula (1) model

For example, formula (1) and formula (2) can be used to express the factors x _t,a that affect the evaluation result of the recommended item information:

among them,

Is a parameter item, used to indicate user preferences; b _{a is a} non-parameter item, used to indicate the recommended item information such as the quality of the material itself;

Represents the optimal model parameters, which are used to describe the real preferences of users; x _t,a represents the feature information of candidate a in the candidate data set to be recommended when it is recommended for the tth time; n represents the total number of experimental rounds; L represents Candidate set size.

Fig. 1 is a schematic flowchart of an embodiment of an application information recommendation method, as shown in Fig. 1, including:

Step 101: Update the parameter items and non-parameter items of the evaluation model according to the user behavior feedback information of the recommended candidates, and generate user characteristic information that represents the user's preferences based on the user attribute information and the user behavior feedback information.

In an exemplary embodiment, for the same user, when the user recommends information for the first time, this step further includes:

Use the initialized evaluation model to evaluate each candidate in the candidate data set to be recommended, and recommend the recommended candidate according to the evaluation result; or,

Use any method in the related art to recommend the recommended candidate to the user from the candidate data set to be recommended.

In an exemplary embodiment, the user attribute information may include, but is not limited to, the user's age, the user's gender, and the province or city where the user is located.

In an exemplary instance, the user behavior feedback information may include, but is not limited to, for example: clicking on the recommended candidate or not clicking on the recommended candidate, or agreeing to use the recommended candidate or disagreeing to use the recommended candidate, etc.

In an exemplary embodiment, updating the parameter items of the evaluation model according to user behavior feedback information may include:

Update the parameter matrix based on the parameter matrix of the recommended candidate (hereinafter referred to as the current) and the current recommended candidate e (we use a to represent any candidate, and e to represent the recommended candidate). A;

The new parameter vector μ is determined according to the updated parameter matrix, the feature information of the currently recommended candidate e and the user behavior feedback information.

In an exemplary embodiment, the following formula (3) to formula (6) can be used to express:

A _t+1 ←A _t +Δx _t,e Δ ^T x _t,e (5)

μ _t+1 ←μ _t +Δx _t,e Δr _t,e (6)

Among them, A _{t+1 in} formula (5) and μ _t+1 in formula (6) respectively represent the latest parameter matrix after update and the latest parameter vector after update; x _t in formula (3) _{, e} represents the feature vector of the recommended candidate _{e at the} current time t (the recommended candidate e is a certain candidate a in the candidate data set to be recommended),

Represents the mean value of the feature vector of the recommended candidate e; r _t,e in formula (4) represents the user's user behavior feedback information on the recommended candidate e. Taking click as an example, if the user clicks on the recommended candidate e, Then, r _t,e is 1. If the user does not click the recommended candidate e, then r _t,e is 0;

Indicates the average value of user behavior feedback information of the user to the recommended candidate e,

Is the mean value of the feature vector of the recommended candidate e (which can be calculated by the following formulas (8) to (9)).

In this way, through formula (7), the current model parameter θ _{t+1 is} obtained by using the updated parameter matrix and the updated parameter vector.

In an exemplary embodiment, updating the non-parameter items of the model according to user behavior feedback information may include, for example:

Determine the new non-parametric estimate of the recommended candidate e according to the number of recommendations of the current recommended candidate e, the updated number of recommendations, and the non-parametric estimated value of the recommended candidate e;

Determine the new feature vector mean value of the recommended candidate e according to the number of recommendations of the currently recommended candidate e, the updated number of recommendations, the average feature vector of the currently recommended candidate e and the feature information of the currently recommended candidate e .

In an illustrative example, the following calculation formula can be used to express:

T _t+1 (e)←T _t (e)+1;

Among them, T _t (e) is the recommended number of recommended candidates e at the current time t,

The mean value of the latest user behavior feedback for the recommended candidate e,

Is the latest feature vector mean value of the recommended candidate e.

In an exemplary embodiment, generating user characteristic information that represents user preferences may include:

For each candidate a in the candidate data set S to be recommended, generate feature information x _t,a of candidate a at the current time t to construct a feature set

This application particularly emphasizes that the required feature information meets at least the following two conditions:

The intersection feature between candidate a and the user, that is, the user attribute information and the feature of candidate a (including but not limited to the attributes of candidate a, such as the content information of the picture (video), etc.) are crossed, and the intersection is obtained, and the cross feature is obtained as Indicates user characteristic information that the user likes. This type of feature is mainly used to represent the intersection information between the user and the candidate data;

The user's own characteristics are mainly used to indicate the user's preferences, as user characteristic information that indicates the user's preferences.

It should be noted that the actual application may not be limited to the above two types of features.

This application emphasizes that the feature information needs to meet at least the above two conditions. However, any feature generation method can be applied to how to generate it, and the specific implementation method is not used to limit the protection scope of this application, and will not be repeated here.

This application uses user behavior feedback information to construct various types of user characteristic information, achieving the purpose of portraying user preferences. Moreover, by constructing rich user characteristics, the subsequent estimation result of the parameter part is better, so that the result returned by the whole system has a higher quality.

In an exemplary embodiment, before this step, the method further includes: obtaining the candidate data set S to be recommended from the background database.

In an exemplary embodiment, if the candidate data set S to be recommended is very large, it may further include:

The candidate data set S to be recommended is sorted to reduce the size of the entire candidate data set S to be recommended, thereby reducing the time-consuming subsequent processing for the candidate data set S to be recommended, and improving the service performance of information recommendation.

In an exemplary embodiment, the sorting process may adopt coarse sorting, etc., which may be a very lightweight sorting process of scoring. This process of counting and sorting can use a very simple value estimation model to estimate each candidate data. In this way, the Top-K candidate data with the highest order of the estimation result can be formed into the candidate data set S to be recommended.

Step 102: When information recommendation is made to the user again, evaluate each candidate in the candidate recommendation data set according to the updated evaluation model and user characteristic information, recommend candidates according to the evaluation result, and return to step 101.

In an exemplary embodiment, evaluating each candidate in the candidate data set to be recommended according to the updated evaluation model and user characteristic information in this step, and recommending the candidates according to the evaluation result, may include:

Taking the updated parameter items and non-parameter items, and the generated user characteristic information as the input of the evaluation model, using the semi-parametric gambling machine to score each candidate in the recommended candidate data set to obtain the θ _t in formula (2) ^* And sort, and recommend the candidate with the highest score, that is, the best, to the user as the candidate when the user is again recommended for information.

In an exemplary example, the candidates in the candidate data set to be recommended are scored and sorted, and the candidate with the highest score is used as the candidate when the user is again recommended for information, including:

For each candidate in the candidate data set to be recommended, calculate the estimate of each candidate according to the model parameter vector information, the feature vector information of the candidate, the user behavior feedback information of the candidate, and the average user behavior feedback Value P _t (a);

The estimated value of each candidate set P _t (a), for each of the candidates are sorted, the maximum value of the estimated candidates to the user as a _t again when information recommendation candidates recommended to the user.

In an exemplary example, using a semi-parameter gambling machine to score and sort the candidates in the candidate data set S to be recommended may include:

First, for each candidate a in the candidate data set S to be recommended, the following calculation is performed to obtain the estimated value P _t (a) of each candidate as shown in formula (10):

Among them, P _t (a) is the estimated value of each candidate; x _t,a represents the feature vector of candidate a at the current time t,

Represents the mean value of the feature vector of candidate a, T _t (a) represents the number of times t candidate a is recommended at the current moment,

Represents the inverse matrix of the parameter matrix, θ _t represents the model parameter vector,

Indicates the average value of user behavior feedback of users to candidate a, and γ _t (a) represents user behavior feedback information of users to candidate a.

T _t (a),

θ _t is obtained by updating the evaluation model. For details on how to obtain it, please refer to the update process of parameter items and non-parameter items in step 101 above, which will not be repeated here.

In the process of estimating the value of candidate data, this application combines the advantages of parameters and non-parameters, that is, it achieves rapid convergence and ensures that the theoretical loss distance from the optimal solution is short, thereby ensuring that the resulting solution is instantaneous The recommendation results are of high quality, which improves the user experience.

Then, according to the estimated value P _t (a) of each candidate set, the candidates are sorted, and the candidate a with the largest estimated value is recommended to the user.

In an illustrative example, if the candidate with the highest estimated value includes two or more than two, one of the candidates can be randomly selected and recommended to the user, or the highest estimated value can be selected after another evaluation based on the non-parameters One of the candidate items of is recommended to the user, and the highest candidate can be selected and recommended to the user after another evaluation based on the parameter items. Of course, it is also possible to add weights to the parameter items and non-parameter items before performing the evaluation in step 102 Then select the candidate with the highest evaluation and recommend it to the user, and so on.

The method provided by this application that uses a semi-parameter gambling machine to score and sort the candidates in the recommended candidate data set S realizes the dynamic capture of user interest information, which is especially suitable for scenarios where user interests change rapidly or lack users Cold start of behavioral data.

This application introduces non-parametric terms through a semi-parametric environment, so that the evaluation model can fit the distribution of unknown data. In the process of estimating the value of the candidate data, this application combines the advantages of the parametric model and the non-parametric model, that is, it achieves rapid convergence and ensures that the theoretical loss distance from the optimal solution is short, thereby ensuring the generated The solution is the recommended result of high quality.

In an exemplary embodiment, before step 101, the information recommendation method of this application further includes:

Step 100: Obtain user behavior data, and process the obtained user behavior data to obtain user attribute information and user behavior feedback information.

In an exemplary example, user behavior data is user behavior feedback information, including but not limited to, click on the recommended candidate or not click on the recommended candidate, or agree to use the recommended candidate or disagree to use the recommended candidate Candidates, etc.

In an illustrative example, processing the obtained user behavior data may include:

Perform corresponding format analysis on the obtained user behavior data, so as to parse the obtained user behavior data into a data format supported by subsequent processing such as step 101.

In an illustrative example, obtaining user behavior data may include:

Obtain feedback user behavior data through the user's terminal equipment such as mobile phones.

In an exemplary embodiment, before obtaining user behavior data, it may further include:

Determine whether the user has feedback. If the user has behavior feedback, continue to perform the step of obtaining user behavior data; if the user has no behavior feedback, then directly end this process.

In this step, user behavior data is collected in real time, and the user behavior data is preprocessed, which facilitates subsequent further processing. In this application, real-time data processing realizes the timely capture of changes in the user's points of interest, thereby more accurately obtaining the user's preferences that change with time and environment, and thus better guarantees the improvement of information recommendation performance.

The present application also provides a computer-readable storage medium that stores computer-executable instructions, and the computer-executable instructions are used to execute any one of the above information recommendation methods.

The present application further provides a device for implementing information recommendation, including a memory and a processor, wherein the memory stores the following instructions that can be executed by the processor for executing the steps of any of the information recommendation methods described above.

Figure 2 is a schematic diagram of the composition structure of an embodiment of the information recommendation device of this application. As shown in Figure 2, it at least includes: a training module, a generating module, and an evaluation module; wherein,

The generating module is used for each candidate in the candidate data set S to be recommended, and generate user characteristic information that represents the user's preference based on user attribute information and user behavior feedback information of the user on the recommended candidate;

In an exemplary embodiment, the information recommendation device of the present application further includes:

The collection module is used to obtain user behavior data, and process the obtained user behavior data to obtain user attribute information and user behavior feedback information.

In an exemplary embodiment, the information recommendation module of this application further includes:

The recall module is used to obtain the candidate data set S to be recommended from the database.

In an exemplary embodiment, the recall module is also used to: sort the candidate data set S to be recommended to reduce the size of the entire candidate data set S to be recommended.

Although the implementation manners disclosed in this application are as described above, the content described is only the implementation manners used for facilitating the understanding of the application, and is not intended to limit the application. Any person skilled in the field to which this application belongs, without departing from the spirit and scope disclosed in this application, can make any modifications and changes in the implementation form and details, but the patent protection scope of this application still requires The scope defined by the appended claims shall prevail.

Claims

An information recommendation method, including:

Update the parameter items and non-parameter items of the evaluation model according to the user behavior feedback information of the recommended candidates, and generate user characteristic information that represents the user's preferences based on the user attribute information and user behavior feedback information;

When information recommendation is made to the user again, each candidate in the candidate data set to be recommended is evaluated according to the updated evaluation model and user characteristic information, the candidate is recommended according to the evaluation result, and the update step is returned.
The information recommendation method according to claim 1, before the method further comprises:

The candidate data set to be recommended is sorted to reduce the size of the candidate data set to be recommended.
The information recommendation method according to claim 1 or 2, before the method further comprising:

Recommending the recommended candidate to the user from the candidate data set to be recommended.
The information recommendation method according to claim 3, wherein said updating the parameter items of the evaluation model according to user behavior feedback information comprises:

Updating the parameter matrix according to the parameter matrix when the recommended candidate is recommended and the characteristic information of the recommended candidate;

A new parameter vector is determined according to the updated parameter matrix, the characteristic information of the recommended candidate and the user behavior feedback information.
The information recommendation method according to claim 3, wherein said updating the non-parameter items of the model according to user behavior feedback information comprises:

Determine the new non-parametric estimated value of the recommended candidate according to the recommended times of the recommended candidate, the updated recommended times and the non-parametric estimated value of the recommended candidate;

Determine the new feature vector of the recommended candidate according to the number of recommendations of the recommended candidate, the updated number of recommendations, the mean value of the feature vector of the recommended candidate, and the feature information of the recommended candidate Mean.
The information recommendation method according to claim 1 or 2, wherein said generating user characteristic information representing user preferences includes:

For each candidate item in the candidate data set to be recommended, feature information of the candidate item is generated according to the intersection of the user attribute information and the feature of the candidate item to construct the feature set representing user feature information that the user likes .
The information recommendation method according to claim 1 or 2, wherein the evaluating each candidate in the candidate data set to be recommended according to the updated evaluation model and user characteristic information, and recommending the candidates according to the evaluation result, includes:

The updated parameter items, the non-parameter items, and the generated user characteristic information are used as the input of the evaluation model, and the candidates in the candidate data set to be recommended are scored and sorted, and the one with the highest score The candidate items are recommended to the user as the candidates when the information is recommended to the user again.
8. The information recommendation method according to claim 7, wherein the candidates in the candidate data set to be recommended are scored and sorted, and the candidate with the highest score is used as the information recommendation for the user again The candidates include:

For each candidate in the candidate data set to be recommended, the value of each candidate is calculated according to the model parameter vector information, the feature vector information of the candidate, the user behavior feedback information for the candidate, and the average user behavior feedback. Estimated value P t (a);

The estimated value of each candidate set P t (a), for each of the candidates are sorted, the maximum value of the estimated candidates as a t the candidate again when information recommendation to the recommendation to the user述Users.
A computer-readable storage medium storing computer-executable instructions for executing the information recommendation method according to any one of claims 1 to 8.
A device for realizing named entity recognition, comprising a memory and a processor, wherein the memory stores the following instructions that can be executed by the processor: for executing the information recommendation of any one of claims 1 to 8 Method steps.
An information recommendation device includes: a training module, a generation module, and an evaluation module; wherein,

The training module is used to update the parameter items and non-parameter items of the evaluation model according to the user behavior feedback information of the recommended candidates;

The generating module is used for each candidate in the candidate data set to be recommended, and generate user characteristic information that represents the user's preferences based on user attribute information and user behavior feedback information of the user on the recommended candidate;

The evaluation module is used to evaluate each candidate in the candidate data set to be recommended according to the updated evaluation model and user characteristic information when information is recommended to the user again, and recommend the candidates according to the evaluation result.
The information recommendation device according to claim 11, further comprising:

The collection module is used to obtain user behavior data, and process the obtained user behavior data to obtain the user attribute information and the user behavior feedback information.
The information recommendation device according to claim 11, further comprising:

The recall module is used to obtain the candidate data set to be recommended from the database.
According to the information recommendation device according to claim 13, the recall module is further configured to:

The candidate data set to be recommended is sorted to reduce the size of the candidate data set to be recommended.