Behavior prediction model training method, behavior prediction method and device, and equipment

Abstract

The embodiments of the present disclosure provide a method for training a behavior prediction model, a behavior prediction method, a device, an electronic device, and a computer-readable storage medium, and relate to the field of computer technology. The method for training the behavior prediction model includes: obtaining the behavior data of multiple user groups on target information, and generating a label group of the target information based on the behavior data; inputting the characteristic information of the target information into the behavior prediction model to be trained to obtain a user behavior prediction data group of the target information; determining the loss function of the behavior prediction model based on the label group, the user behavior prediction data group, and the data distribution state in the user behavior prediction data group; adjusting the model parameters of the behavior prediction model based on the loss function to train the behavior prediction model. The technical solution of the embodiments of the present disclosure can improve the accuracy of the model.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of computer technology. More specifically, the embodiments of the present disclosure relate to a training method and device for a behavior prediction model, a behavior prediction method and device, an electronic device, and a computer-readable storage medium. Background Art

[0002] This section is intended to provide a background or context to the embodiments of the disclosure that are recited in the claims. No description herein is admitted to be prior art by inclusion in this section.

[0003] With the development of big data, understanding users’ reading preferences is crucial to the accuracy of recommendation results. Therefore, it is necessary to accurately determine the user behavior tendencies of all users for a certain target information.

[0004] In related technologies, when analyzing user groups for user behaviors related to target information, it is generally calculated through heuristic algorithms and supervised learning algorithms. Among them, heuristic algorithms manually read and investigate a large amount of data, summarize the rules from the large amount of data, and then design corresponding rule engines to complete the analysis of user groups. The supervised learning-based algorithm uses a large amount of manually labeled data to learn the mapping relationship between target information and user groups. Summary of the invention

[0005] However, in the related art, the summarized rule engines are difficult to cover all situations, have certain limitations and are difficult to reuse, and the accuracy of the mapping relationship learned through manually labeled data is low.

[0006] Therefore, there is a great need for an improved method for training a behavior prediction model so as to accurately predict the user behavior of each user group based on the trained behavior prediction model.

[0007] In this context, embodiments of the present disclosure are intended to provide a method for training a behavior prediction model, a device for training a behavior prediction model, a behavior prediction method, a behavior prediction device, an electronic device, and a computer-readable storage medium.

[0008] According to one aspect of the present disclosure, a method for training a behavior prediction model is provided, comprising: obtaining behavior data of multiple user groups on target information, and generating a label group for the target information based on the behavior data, the label group including labels corresponding to each user group; inputting feature information of the target information into the behavior prediction model to be trained to obtain a user behavior prediction data group for the target information, the user behavior prediction data group including user behavior prediction data corresponding to each user group; determining a loss function of the behavior prediction model based on the label group, the user behavior prediction data group and data distribution status in the user behavior prediction data group; and adjusting model parameters of the behavior prediction model based on the loss function to train the behavior prediction model.

[0009] In an exemplary embodiment of the present disclosure, generating the tag group of the target information based on the behavior data includes: normalizing the behavior data of the target information by the multiple user groups to obtain normalized behavior data; and generating the tag group of the target information based on the normalized behavior data.

[0010] In an exemplary embodiment of the present disclosure, normalizing the behavior data of the multiple user groups for the target information to obtain normalized behavior data includes: determining a global behavior data maximum value and a global behavior data minimum value in the behavior data of each user group for each sample information; and normalizing the behavior data of each user group for the target information according to the global behavior data maximum value and the global behavior data minimum value to obtain the normalized behavior data.

[0011] In an exemplary embodiment of the present disclosure, after normalizing the behavior data of each user group for the target information according to the global behavior data maximum value and the global behavior data minimum value to obtain the normalized behavior data, the method further includes: determining the local behavior data maximum value and the local behavior data minimum value in the behavior data of each user group for the target information; and re-normalizing the normalized behavior data according to the local behavior data maximum value and the local behavior data minimum value to update the normalized behavior data.

[0012] In an exemplary embodiment of the present disclosure, the loss function of the behavior prediction model is determined based on the label group, the user behavior prediction data group, and the data distribution status in the user behavior prediction data group, including: determining an expected loss term based on the label group and the user behavior prediction data group; determining a regularization term based on the data distribution status of multiple user groups in the user behavior prediction data group; and logically combining the expected loss term and the regularization term to determine the loss function.

[0013] In an exemplary embodiment of the present disclosure, the regularization term includes a first regularization term and a second regularization term; the regularization term is determined according to the data distribution status of multiple user groups in the user behavior prediction data group, including: determining the first regularization term according to the distribution status of the prediction data of adjacent user groups in the multiple user groups; determining the second regularization term according to the prediction data of the target user group in the multiple user groups; the target user group is the user group corresponding to the label with the largest value in the label group.

[0014] In an exemplary embodiment of the present disclosure, determining the first regularization term according to the distribution status of the prediction data of adjacent user groups among the multiple user groups includes: calculating the difference between the prediction data of the adjacent user groups, and determining the first regularization term according to the difference.

[0015] According to one aspect of the present disclosure, a behavior prediction method is provided, comprising: obtaining information to be processed, and obtaining multiple feature information of the information to be processed; inputting the multiple feature information into a behavior prediction model, predicting user behaviors of multiple user groups with respect to the information to be processed, and obtaining prediction data corresponding to the multiple user groups; wherein the behavior prediction model is a model trained according to the training method of the behavior prediction model described in any one of the above items.

[0016] According to one aspect of the present disclosure, a training device for a behavior prediction model is provided, comprising: a label generation module, used to obtain behavior data of multiple user groups on target information, and generate a label group for the target information based on the behavior data, the label group including labels corresponding to each user group; a prediction data determination module, used to input feature information of the target information into the behavior prediction model to be trained, and obtain a user behavior prediction data group for the target information, the user behavior prediction data group including user behavior prediction data corresponding to each user group; a loss function determination module, used to determine the loss function of the behavior prediction model based on the label group, the user behavior prediction data group and the data distribution status in the user behavior prediction data group; and a model training module, used to adjust model parameters of the behavior prediction model according to the loss function to train the behavior prediction model.

[0017] In an exemplary embodiment of the present disclosure, the label generation module includes: a normalization module, used to normalize the behavior data of the multiple user groups on the target information to obtain normalized behavior data; and a generation control module, used to generate a label group for the target information based on the normalized behavior data.

[0018] In an exemplary embodiment of the present disclosure, the normalization module includes: a global data determination module, which is used to determine the global behavior data maximum value and the global behavior data minimum value in the behavior data of each user group for each sample information; a normalization control module, which is used to normalize the behavior data of each user group for the target information according to the global behavior data maximum value and the global behavior data minimum value to obtain the normalized behavior data.

[0019] In an exemplary embodiment of the present disclosure, after normalizing the behavior data of each user group for the target information according to the global behavior data maximum value and the global behavior data minimum value to obtain the normalized behavior data, the device also includes: a local data determination module, used to determine the local behavior data maximum value and the local behavior data minimum value in the behavior data of each user group for the target information; a normalization update module, used to re-normalize the normalized behavior data according to the local behavior data maximum value and the local behavior data minimum value to update the normalized behavior data.

[0020] In an exemplary embodiment of the present disclosure, the loss function determination module includes: an expected loss term determination module, which is used to determine the expected loss term based on the label group and the user behavior prediction data group; a regularization term determination module, which is used to determine the regularization term based on the data distribution status of multiple user groups in the prediction data group; and a logical combination module, which is used to logically combine the expected loss term and the regularization term to determine the loss function.

[0021] In an exemplary embodiment of the present disclosure, the regularization term includes a first regularization term and a second regularization term; the regularization term determination module includes: a first determination module, used to determine the first regularization term according to the distribution state of the prediction data of the adjacent user groups among the multiple user groups; a second determination module, used to determine the second regularization term according to the prediction data of the target user group among the multiple user groups; the target user group is the user group corresponding to the label with the largest value in the label group.

[0022] In an exemplary embodiment of the present disclosure, the first determination module includes: a difference calculation module, configured to calculate the difference between the prediction data of the adjacent user groups, and determine the first regularization term according to the difference.

[0023] According to one aspect of the present disclosure, a behavior prediction device is provided, comprising: a feature information acquisition module, used to acquire information to be processed, and acquire multiple feature information of the information to be processed; a prediction module, used to input the multiple feature information into a behavior prediction model, predict user behaviors of multiple user groups with respect to the information to be processed, and obtain prediction data corresponding to the multiple user groups; wherein the behavior prediction model is a model trained according to the training method of the behavior prediction model described in any one of the above items.

[0024] According to one aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing executable instructions; wherein the processor is configured to execute any one of the above-described training methods for a behavior prediction model or any one of the above-described behavior prediction methods by executing the executable instructions.

[0025] According to one aspect of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the training method of the behavior prediction model as described in any one of the above items or the behavior prediction method as described in any one of the above items is implemented.

[0026] According to the training method and device of the behavior prediction model, the behavior prediction method and device, the electronic device and the computer-readable storage medium of the embodiment of the present disclosure, an accurate label group can be generated according to the behavior data of the user group on the target information, and the modeling between the target information and the user group is completed through the characteristic information of the target information and the user behavior prediction data group. In addition, the loss function of the behavior prediction model is determined according to the label group, the user behavior prediction data group and the data distribution state in the user behavior prediction data group. This enables the behavior prediction model to cover all situations, avoid limitations and enable reuse. It can make the established behavior prediction model more in line with practical applications from multiple dimensions, thereby improving the accuracy and reliability of the behavior prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, in which:

[0028] Figure 1 A schematic diagram of a system architecture of an application scenario of an embodiment of the present disclosure is schematically shown.

[0029] Figure 2 The flowchart of the training method of the behavior prediction model in the embodiment of the present disclosure is schematically shown.

[0030] Figure 3The following is a schematic diagram of the process of determining a tag group in an embodiment of the present disclosure.

[0031] Figure 4 The schematic diagram of the architecture of the behavior prediction model in the embodiment of the present disclosure is schematically shown.

[0032] Figure 5 The flowchart of determining the loss function in an embodiment of the present disclosure is schematically shown.

[0033] Figure 6 The flowchart of the behavior prediction method according to the embodiment of the present disclosure is schematically shown.

[0034] Figure 7 A schematic block diagram of a training device for a behavior prediction model according to an embodiment of the present disclosure is schematically shown.

[0035] Figure 8 A schematic block diagram of a behavior prediction device according to an embodiment of the present disclosure is schematically shown.

[0036] Fig. 9 A block diagram of an electronic device according to an embodiment of the present disclosure is schematically shown.

[0037] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts. DETAILED DESCRIPTION

[0038] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0039] Those skilled in the art will appreciate that the embodiments of the present disclosure may be implemented as a system, device, apparatus, method or computer program product. Therefore, the present disclosure may be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0040] According to an embodiment of the present disclosure, a training method for a behavior prediction model, a training device for a behavior prediction model, a behavior prediction method, a behavior prediction device, an electronic device, and a computer-readable storage medium are proposed.

[0041] In addition, any number of elements in the drawings is for illustration and not limitation, and any naming is only for distinction and does not have any limiting meaning.

[0042] The principle and spirit of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure. SUMMARY OF THE INVENTION

[0044] The inventors found that for heuristic algorithms, the summarized rule engines are often difficult to cover all situations, so such algorithms often do not have good generalization performance, and the algorithm flow based on heuristics is difficult to reuse and consumes a lot of manpower. Algorithms based on supervised learning are difficult to obtain high-quality labeled data, so the accuracy is poor.

[0045] Based on the above content, the basic idea of ​​the present disclosure is to replace manually designed labels with user behavior data in the system as the basis for modeling and predicting user behavior.

[0046] More specifically, in the present disclosure, the behavior data of multiple user groups on target information are obtained, and a label group of the target information is generated based on the behavior data, and the label group includes labels corresponding to each user group; the characteristic information of the target information is input into the behavior prediction model to be trained to obtain a user behavior prediction data group of the target information, and the user behavior prediction data group includes user behavior prediction data corresponding to each user group; based on the label group, the user behavior prediction data group, and the data distribution state in the user behavior prediction data group, the loss function of the behavior prediction model is determined; according to the loss function, the model parameters of the behavior prediction model are adjusted to train the behavior prediction model. Therefore, according to the embodiments of the present disclosure, the accuracy and reliability of the training of the behavior prediction model can be improved.

[0047] After introducing the basic principles of the present disclosure, various non-limiting embodiments of the present disclosure are described in detail below.

[0048] Application Scenario Overview

[0049] It should be noted that the following application scenarios are only shown to facilitate understanding of the spirit and principle of the present disclosure, and the embodiments of the present disclosure are not limited in this respect. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0050] First reference Figure 1 , Figure 1 The system architecture diagram of an exemplary application scenario of an embodiment of the present disclosure is shown. The training method of the behavior prediction model can be used in application scenarios such as browsing products, browsing articles, and browsing videos. Figure 1As shown, the system architecture 100 includes a terminal 101, a network 102 and a server 103. The server 103 analyzes the average click rate of multiple user groups extracted from the terminal 101 on the target information, so that the server predicts the tendency of user behavior of each user group on the target information, and then recommends appropriate content to each user group based on the tendency. Figure 1 The schematic framework shown is only an example in which the embodiments of the present disclosure can be implemented. The scope of application of the embodiments of the present disclosure is not limited in any aspect by the framework.

[0051] It should be noted that the server 103 can be a local server or a remote server. In addition, the server 103 can also be other products that can provide storage functions or processing functions, such as cloud servers. The embodiments of the present disclosure are not specifically limited here. The server can also be a terminal device with fast computing capabilities or a vehicle-mounted device, etc., which is not limited here. The terminal 101 can be any device that can generate user behavior data, such as a smart phone, a tablet computer, and a computer.

[0052] It should be understood that in the application scenario of the present disclosure, the actions of the embodiments of the present disclosure can be executed by the server 103 or by a terminal with computing capabilities, and the present disclosure is not subject to any restrictions on the execution subject, as long as the actions disclosed in the embodiments of the present disclosure are executed. Here, the execution subject is taken as an example for specific description.

[0053] Exemplary Methods

[0054] Combine the following Figure 1 For application scenarios, refer to Figure 2 To describe the training method of the behavior prediction model according to the exemplary embodiment of the present disclosure. It should be noted that the above application scenarios are only shown to facilitate understanding of the spirit and principles of the present disclosure, and the embodiments of the present disclosure are not limited in this regard. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0055] Figure 2 FIG. 1 is a flow chart showing a method for training a behavior prediction model according to an embodiment of the present disclosure. Figure 2 As shown, the training method of the behavior prediction model may include the following steps:

[0056] In step S210, behavior data of multiple user groups on target information is obtained, and a tag group of the target information is generated according to the behavior data, wherein the tag group includes tags corresponding to each user group;

[0057] In step S220, the characteristic information of the target information is input into the behavior prediction model to be trained to obtain a user behavior prediction data group of the target information, wherein the user behavior prediction data group includes user behavior prediction data corresponding to each user group;

[0058] In step S230, a loss function of the behavior prediction model is determined according to the tag group, the user behavior prediction data group, and the data distribution state in the user behavior prediction data group;

[0059] In step S240, the model parameters of the behavior prediction model are adjusted according to the loss function to train the behavior prediction model.

[0060] The technical solution in the disclosed embodiments completes the modeling between the target information and the user group through the feature information of the target information and the user behavior prediction data group, and determines the loss function of the behavior prediction model according to the label group, the user behavior prediction data group and the data distribution status in the user behavior prediction data group. This enables the behavior prediction model to cover all situations, avoids limitations and can be reused in multiple scenarios, making the behavior prediction model more in line with practical applications and improving the accuracy and reliability of the behavior prediction model.

[0061] Next, the training method of the behavior prediction model in the embodiment of the present disclosure is explained in detail with reference to the accompanying drawings.

[0062] In step S210, behavior data of multiple user groups on target information is obtained, and a tag group of the target information is generated according to the behavior data, wherein the tag group includes tags corresponding to each user group.

[0063] In the disclosed embodiment, the target information may be any type of information on the target platform. The target platform may be a news platform, an interactive platform, a shopping platform, etc., and a news platform is used as an example for explanation herein. The target information may be text information or image information, etc., and text information is used as an example for explanation herein. When the target platform is a news platform, the target information may be, for example, any article on the news platform, such as article 1. Multiple user groups refer to groups determined by dividing all users according to age distribution, and the specific number of divided user groups may be determined according to actual needs.

[0064] For example, a group of seed users with age information on a news platform are selected, and the users are divided into n user groups according to their age distribution. Among them, any user can be represented by u, and The age group is represented by age i user groups.

[0065] Behavioral data refers to the data generated by the user behavior of each user group on the target information. The user behavior here can be reading behavior, which can be expressed as click, browse, collect, forward, comment and other operation behaviors. Here, the user behavior is explained by the click operation. Behavioral data can be used to describe the behavioral characteristics of each user group on the target information, and the behavioral data of each user group can be the same or different. In the embodiment of the present disclosure, for each piece of target information, the behavioral data of each user group with respect to the target information can be obtained, and the behavioral data can be expressed by the average click-through rate. The average click-through rate refers to the average of the click-through rates of each user group on the target information. In the embodiment of the present disclosure, the average click-through rate of multiple user groups on the target information can be expressed as follows: in The age group is represented by age i User groups The average click-through rate of the target information. When the target information (a certain article) is fully exposed to all age groups, the reading tendency of the target information in each age group is approximately equal to the average click-through rate of the target information in each age group. Click-through rate refers to the number of times a specified content is clicked and exposed on a website or application platform. Click-through rate is usually an important measurement indicator in a recommendation system. Therefore, this behavior data can be used to describe the reading tendency or reading tendency value of a user group.

[0066] In the disclosed embodiment, for the target information, multiple feature information and behavior data can be determined. The multiple feature information can be used to describe the content of the target information, and the multiple feature information can include but is not limited to titles, keywords, and categories. The feature information and behavior data can be further combined in a fixed format.

[0067] For example, for the target information doc i Its feature information and behavior data are organized into the following data format, expressed as: in, Target information doc i Title, w i is the i-th word in the title, and t is the length of the title. Target information doc i A set of keywords, where word i is the i-th keyword of the target information, and k is the number of keywords. Is the target information doc i Category. Age j The user groups in the target information doc i The average click-through rate above.

[0068] After obtaining the behavior data of multiple user groups on the target information, a label group can be generated based on the behavior data of multiple user groups. The label group can include labels corresponding to each user group. The labels here can be used to represent the fitting object of the behavior prediction model, that is, the target to be output. Since no manual labeling is performed, the generated label group can be used as pseudo-labeled data to improve the accuracy of the labeling, so as to facilitate the processing of the model.

[0069] Figure 3 A flowchart for determining a tag group is schematically shown in FIG. Figure 3 As shown in , it mainly includes the following steps:

[0070] In step S310, the behavior data of the multiple user groups on the target information are normalized to obtain normalized behavior data.

[0071] In this step, the click-through rate data of user groups in some age groups on the target information is relatively high, while the click-through rate data of user groups in other age groups on the target information is relatively low. In order to avoid the influence of factors such as inconsistent behavioral data benchmarks of user groups in different age groups, the behavioral data can be normalized multiple times to obtain normalized behavioral data. Specifically, the behavioral data of user groups in the same age group on multiple sample information (different articles) can be normalized to achieve normalization on age groups, so that the behavioral data of user groups in different age groups on the same sample information are comparable, which can be closer to the reading tendencies of user groups in different age groups, and improve the accuracy and authenticity of the behavioral data.

[0072] The specific steps of normalizing the behavior data of multiple user groups on the target information include: determining the global behavior data maximum value and the global behavior data minimum value in the behavior data of each user group on each sample information; normalizing the behavior data of each user group on the target information according to the global behavior data maximum value and the global behavior data minimum value to obtain the normalized behavior data. Wherein, each sample information refers to all information of the target platform, such as Article 1, Article 2, and Article 3, etc. The target information can be any one of the multiple sample information. The global behavior data maximum value refers to the behavior data with the largest value in the behavior data of each user group on multiple sample information, such as the maximum value determined from the average click-through rate of each user group on Article 1, Article 2, and Article 3. The global behavior data minimum value refers to the behavior data with the smallest value in the behavior data of each user group on multiple sample information, such as the minimum value determined from the average click-through rate of each user group on Article 1, Article 2, and Article 3. Furthermore, the behavior data of each user group on the target information can be normalized according to the global behavior data maximum value and global behavior data minimum value, so as to obtain normalized behavior data, that is, the average click rate of each user group on the target information is normalized to obtain the normalized click rate. The normalization here refers to the maximum and minimum value normalization, which can be specifically expressed by formula (1):

[0073]

[0074] in, Behavioral data for each user group on the target information, i.e., average click-through rate, is the minimum value of global behavior data, It is the maximum value of global behavior data.

[0075] By normalizing the behavioral data, the average click rate of the user group can be made closer to the actual reading tendency of the user group in that age group, thereby improving the accuracy of the behavioral data.

[0076] After obtaining the normalized behavior data, the absolute values ​​of the behavior data of articles with similar topics and semantics may be inconsistent due to factors in the recommendation system. Therefore, the normalized behavior data normalized on the age group is normalized again for the target information (inside the article) to update the normalized behavior data.

[0077] When updating the normalized behavior data, the local behavior data maximum value and the local behavior data minimum value can be determined in the behavior data of each user group for the target information, and the normalized behavior data can be further normalized again based on the local behavior data maximum value and the local behavior data minimum value to update the normalized behavior data. The local behavior data minimum value refers to the behavior data with the smallest value in the behavior data of each user group for the target information, that is, the minimum value determined from the average click-through rate of each user group for Article 1. The local behavior data maximum value refers to the behavior data with the largest value in the behavior data of each user group for the target information, for example, the maximum value determined from the average click-through rate of each user group for Article 1. Furthermore, the normalized behavior data can be normalized again based on the local behavior data maximum value and the local behavior data minimum value to update the normalized behavior data, and can be normalized again using formula (2):

[0078]

[0079] in, is the minimum value of local behavior data, is the maximum value of local behavior data, Behavioral data for each user group on the targeted information.

[0080] In this way, for the target information, the updated normalized behavior data can be used to analyze the reading tendency of the user groups in each age group towards the target information.

[0081] In step S320, a label group of the target information is generated according to the normalized behavior data.

[0082] In the disclosed embodiment, after obtaining the normalized behavior data, the updated normalized behavior data can be determined as labels corresponding to each user group, so as to form a label group of target information according to the labels corresponding to each user group, and the label group is used to describe the reading tendency of a group of user groups of different age groups. By using the label group as pseudo-annotated data, the fitting object of the model is represented.

[0083] In the disclosed embodiment, pseudo-annotated data between target information and age groups is generated through behavioral data of user groups on target information, and the pseudo-annotated data is used as a label group, which can make the generated label group more accurate and more in line with the actual situation than manually annotated labels.

[0084] In step S220, the feature information of the target information is input into the behavior prediction model to be trained to obtain a user behavior prediction data group of the target information, wherein the user behavior prediction data group includes user behavior prediction data corresponding to each user group.

[0085] In the disclosed embodiments, the characteristic information of the target information includes but is not limited to keywords, titles, and categories. The behavior prediction model to be trained can be various suitable models. Here, the behavior prediction model to be trained is taken as a BERT model as an example for explanation. After obtaining the characteristic information of the target information, the characteristic information can be used as input data and input into the behavior prediction model to be trained, so that the behavior prediction model to be trained processes the characteristic information to obtain a user behavior prediction data group of the target information. It should be noted that the user behavior prediction data group may include user behavior prediction data corresponding to each user group.

[0086] For example, the characteristic information of the target information and the label group can be determined as the training sample of the behavior prediction model to be trained for model training. The format of the training sample can be expressed as The title Keywords and categories It can be used as input for the behavior prediction model to be trained. is the normalized label group of each user group in the target information. The final output result of the behavior prediction model to be trained is the user behavior prediction data group of the target information. The user behavior prediction data group can be expressed as in Represents the target information doc i The behavior prediction model to be trained is for age group j Predictive data on the reading preferences of user groups.

[0087] Figure 4 The schematic diagram shows the architecture of the behavior prediction model to be trained, see Figure 4 As shown in , the feature information of the target information can be used as the input of the behavior prediction model to be trained, and the output result of the model is obtained after passing through the embedding layer and the transmission layer, and the final prediction result is further obtained through the softmax layer. The prediction result is the user behavior prediction data set of the target information. That is, the entire processing process is: converting the title, keywords and categories of the target information into multiple text vectors, and concatenating multiple text vectors and separator vectors to determine the input vector, and making predictions based on the input vector to obtain the user behavior prediction data of multiple user groups in the target information.

[0088] Specifically, refer to Figure 4As shown in the figure, the input layer of the BERT model consists of the title, keywords, and categories of the target information, which are concatenated after passing through the embedding layer. It should be noted that the title, keywords, and categories are all in text form, and the embedding layer of the model converts the feature information into a vector of fixed dimension, that is, multiple text vectors, and there is a one-to-one correspondence between the multiple text vectors and the feature information. Then, the multiple text vectors and the separator vector are concatenated to obtain a vector representation of the target information as the input vector.

[0089] In the embodiment of the present disclosure, the input vector of the behavior prediction model to be trained can be expressed as bert input =concat(E cls ,E title ,E sep ,E cate ,E sep ,E keyword ). Among them, E cls It is a special token vector in the BERT model, which is used to aggregate all the information input by the BERT model, that is, to integrate the title, category, and keyword information. sep It is a marker used to separate different information, that is, a separator vector. title is the title vector in the text vector, E cate is the category vector in the text vector, E keyword is the keyword vector in the text vector, that is, the embedding representation of the title, category, and keyword obtained through the embedding layer of the BERT model. After the input vector passes through the transmission layer, the text vector can be fused to obtain a vector that integrates all the input information, and then the output vector is obtained. Figure 4 As shown in input After passing through the transmission layer of the BERT model, the output vector bert is obtained output =[T cls ,T title ,T sep ,T cate ,T sep ,T keyword ]. Among them, T cls All feature information is integrated, and other vectors correspond one-to-one to the vectors in the input vector.

[0090] Furthermore, a user behavior prediction data set including user behavior prediction data corresponding to each user group can be obtained based on the output vector. That is, the flag vector that integrates all feature information in the output vector can be output to the softmax layer for processing to determine the final prediction result as the user behavior prediction data set. The user behavior prediction data set is used to describe the predicted value of the reading behavior of the target information by the user group of each age group. The user behavior prediction data can be determined by formula (3):

[0091] model predict =softmax(Matrix softmax ×T cls ) Formula (3)

[0092] in, is an n×1 vector, where n is the number of all age groups, and each dimension Representative age group i The reading tendency of the user group. softmax is an n×h matrix, where h is the embedding dimension of the hidden vector in the BERT model, T cls is an h×1 vector, and softmax is a common normalization function in machine learning. Figure 4 As shown in , the predicted value can be any value between 0 and 1.

[0093] In the disclosed embodiment, the user behavior prediction data group of the target information is obtained by using the characteristic information of the target information, which can predict the user behavior tendency of each user group towards the target information from multiple dimensions, improve the accuracy of the user behavior prediction data group, and provide a basis for model training.

[0094] In step S230, a loss function of the behavior prediction model is determined according to the tag group, the user behavior prediction data group, and the data distribution state in the user behavior prediction data group.

[0095] In the disclosed embodiment, the data distribution state in the user behavior prediction data group is used to describe the specific state of the prediction data of each user group in the user behavior prediction data group, such as data change trend, data size, etc. Further, the loss function of the behavior prediction model can be determined based on the tag group and the user behavior prediction data group, as well as the data distribution state in the user behavior prediction data group, so as to construct the behavior prediction model.

[0096] Figure 5 A flowchart for determining the loss function is schematically shown in FIG. Figure 5 As shown in , it mainly includes the following steps:

[0097] In step S510, an expected loss term is determined according to the tag group and the user behavior prediction data group;

[0098] In step S520, a regularization term is determined according to data distribution states of multiple user groups in the user behavior prediction data set;

[0099] In step S530, the expected loss term and the regularization term are logically combined to determine the loss function.

[0100] In the disclosed embodiment, first, the expected loss term can be determined based on the label group and the user behavior prediction data group, and the expected loss term can be used to describe the distribution value of the pseudo-label data. The expected loss term can be calculated using the mean square error. Specifically, the difference between the predicted data of the user group of each age group and the label (updated normalized behavior data) of the user group of that age group in the user behavior prediction data group can be calculated. Specifically, the expected loss term can be calculated using formula (4):

[0101]

[0102] Secondly, in order to make the predictions of the behavior prediction model more in line with objective laws, it is necessary to combine regularization terms to guide the training of the behavior prediction model. Regularization terms are used to describe regularization, which refers to the process of introducing additional information to solve overfitting. In the optimization process of machine learning, regularization terms are often added to the objective function. The regularization terms here may include a first regularization term and a second regularization term, and the types of the first regularization term and the second regularization term are different. The first regularization term is used to describe the changing state of the age group. The second regularization term is used to describe the prediction accuracy. In the embodiment of the present disclosure, the first regularization term and the second regularization term can be determined according to the data distribution status of multiple user groups in the user behavior prediction data group.

[0103] Since age groups change continuously, the differences in preferences of user groups in adjacent age groups are small, while the differences in preferences of user groups in non-adjacent age groups are large. For user groups in each age group, their reading tendency (i.e., predicted data) should be a smoothly changing curve, not a wavy line with violent fluctuations. Therefore, when determining the first regularization term based on the data distribution status of multiple user groups in the user behavior prediction data group, the first regularization term can be determined based on the distribution status of the predicted data of adjacent user groups in the multiple user groups. The adjacent user groups can be any two adjacent user groups. The difference between the predicted data of adjacent user groups in the user behavior prediction data group can be calculated, and then the first regularization term can be determined based on the difference. Specifically, it can be determined by formula (5):

[0104]

[0105] In addition, during the application process, attention will also be paid to whether the predicted data of the user group with the highest tendency in the age group is accurate. Based on this, when determining the second regularization term according to the data distribution status of multiple user groups in the user behavior prediction data group, the second regularization term can be determined according to the evaluation parameters of the predicted data of the target user group in the multiple user groups, and the evaluation parameters can be, for example, accuracy. The target user group refers to the user group corresponding to the label with the largest value in the label group. That is, the age group can be based on the age group corresponding to the maximum value of the normalized behavior data in the updated normalized behavior data. The second regularization term can be a head regularization term to increase the penalty according to the prediction error of the second regularization term. Therefore, the second regularization term can be judged based on the accuracy of the predicted data of the user group with the highest tendency, and can be specifically calculated according to formula (6):

[0106]

[0107] Again, after obtaining the first regularization term and the second regularization term, the expected loss term and the first regularization term and the second regularization term can be logically combined according to the corresponding weights to determine the loss function. The logical combination can be an addition operation. The loss function of the behavior prediction model can be determined according to formula (7):

[0108] loss=0.7*l mse +0.2*l con +0.1*l top Formula (7)

[0109] Among them, the weight of the expected loss term in the loss function is the largest, and the weight corresponding to the weight of the second regularization term is the smallest.

[0110] pass Figure 5 The technical solution in can add a regularization term on the basis of the expected loss term, so that the loss function of the behavior prediction model is more in line with the actual situation and improves the accuracy and reliability of the loss function.

[0111] In step S240, the model parameters of the behavior prediction model are adjusted according to the loss function to train the behavior prediction model.

[0112] In the disclosed embodiment, after determining the loss function of the behavior prediction model, the model parameters of the behavior prediction model can be adjusted according to the loss function to train the behavior prediction model. The loss function indicates the distribution state that the predicted data should conform to, and the label group is the fitting object, that is, the value that is desired to be output. The output of the behavior prediction model is a user behavior prediction data group, which represents the distribution of age group tendencies. The specific process of model training may include: using multiple feature information such as the title, keywords, and categories corresponding to the target information as the input of the model to obtain the predicted data, using the behavior data of multiple user groups on the target information as the label of the model, adjusting the model parameters of the behavior prediction model according to the loss function to perform model training, and fitting the behavior prediction model so that the predicted data obtained by the behavior prediction model is consistent with the label group according to the weight of the loss function.

[0113] The technical solution in the disclosed embodiment can establish a behavior prediction model of the mapping relationship between the text semantics of the target information and the reading tendency of user groups of different age groups, and can determine the loss function through the label group, the user behavior prediction data group and the data distribution state in the user behavior prediction data group, and then train the behavior prediction model based on the loss function to improve the accuracy, reliability and stability of the model. It avoids the limitations of the models in the relevant technology, can be reused, and increases flexibility and application scope. Through the idea of ​​weak supervision learning, using the feature information such as the title, keywords, and category of the target information, as well as the behavior data of user groups of various age groups on the target platform on the target information, the modeling between the target information and the user groups corresponding to the age group is completed, so that the established behavior prediction model can accurately integrate the age group rules and improve the accuracy of the model. Furthermore, two regularization terms, smoothing constraint and head constraint, are added to generate a loss function to train the model, which can make the trained model closer to practical application and more accurate.

[0114] The disclosed embodiments also provide a behavior prediction method, which can be implemented by the behavior prediction model described in the above embodiments. Figure 6 A flow chart of a behavior prediction method is schematically shown in FIG. 1 , and the behavior prediction method can be executed by a server, which can be Figure 1 The server 103 shown in FIG. Figure 6 As shown, the behavior prediction method at least includes steps S610 to S620, which are described in detail as follows:

[0115] In step S610, information to be processed is obtained, and a plurality of feature information of the information to be processed is obtained.

[0116] In the disclosed embodiment, the information to be processed may be of the same type as the target information, such as any article to be processed on a news platform. The multiple feature information of the information to be processed may include, but is not limited to, title, keywords, and category. For example, the information to be processed displayed on the terminal may be obtained, such as article 2 on the target platform. Further, semantic analysis may be performed on article 2 to obtain the title, keywords, and category of the information to be processed.

[0117] In step S620, the plurality of feature information is input into a behavior prediction model, and user behaviors of a plurality of user groups with respect to the information to be processed are predicted to obtain prediction data corresponding to the plurality of user groups.

[0118] In the disclosed embodiment, multiple feature information corresponding to the information to be processed can be input into the behavior prediction model trained according to the above embodiment, so that the behavior prediction model obtains the prediction data of the user behavior of each user group on the information to be processed based on the input feature information. The user behavior here can be reading behavior. The prediction data can be the tendency or degree of tendency of the reading behavior of the user group of each age group for the information to be processed. The behavior prediction method can be applied not only to the estimation of news reading behavior, but also to other aspects, such as the estimation of product clicks when making product recommendations, etc. The tendency can be a value between 0-1, and the tendency is positively correlated with the value. The closer to 1, the higher the tendency of the user group of this age group to read the information to be processed, and vice versa, the closer to 0, the lower the tendency of the user group of this age group to read the information to be processed.

[0119] Furthermore, after determining the reading behavior tendency of each user group towards the processed information, appropriate information can be recommended to each user group based on the prediction results, reducing the number of recommendations that do not match the age group and the recommended information, improving the accuracy and pertinence of the recommendations, and enhancing the user's reading experience and recommendation effect.

[0120] Exemplary Devices

[0121] Next, refer to Figure 7 The training device of the behavior prediction model according to the exemplary embodiment of the present disclosure is described. Figure 7 As shown, the behavior prediction model training device 700 may include:

[0122] The tag generation module 701 is used to obtain the behavior data of multiple user groups on the target information, and generate a tag group of the target information according to the behavior data, wherein the tag group includes tags corresponding to each user group;

[0123] A prediction data determination module 702 is used to input the characteristic information of the target information into the behavior prediction model to be trained to obtain a user behavior prediction data group of the target information, wherein the user behavior prediction data group includes user behavior prediction data corresponding to each user group;

[0124] A loss function determination module 703, used to determine the loss function of the behavior prediction model according to the tag group, the user behavior prediction data group and the data distribution state in the user behavior prediction data group;

[0125] The model training module 704 is used to adjust the model parameters of the behavior prediction model according to the loss function to train the behavior prediction model.

[0126] In an exemplary embodiment of the present disclosure, the label generation module includes: a normalization module, used to normalize the behavior data of the multiple user groups on the target information to obtain normalized behavior data; and a generation control module, used to generate a label group for the target information based on the normalized behavior data.

[0127] In an exemplary embodiment of the present disclosure, the normalization module includes: a global data determination module, which is used to determine the global behavior data maximum value and the global behavior data minimum value in the behavior data of each user group for each sample information; a normalization control module, which is used to normalize the behavior data of each user group for the target information according to the global behavior data maximum value and the global behavior data minimum value to obtain the normalized behavior data.

[0128] In an exemplary embodiment of the present disclosure, after normalizing the behavior data of each user group for the target information according to the global behavior data maximum value and the global behavior data minimum value to obtain the normalized behavior data, the device also includes: a local data determination module, used to determine the local behavior data maximum value and the local behavior data minimum value in the behavior data of each user group for the target information; a normalization update module, used to re-normalize the normalized behavior data according to the local behavior data maximum value and the local behavior data minimum value to update the normalized behavior data.

[0129] In an exemplary embodiment of the present disclosure, the loss function determination module includes: an expected loss term determination module, which is used to determine the expected loss term based on the label group and the user behavior prediction data group; a regularization term determination module, which is used to determine the regularization term based on the data distribution status of multiple user groups in the prediction data group; and a logical combination module, which is used to logically combine the expected loss term and the regularization term to determine the loss function.

[0130] In an exemplary embodiment of the present disclosure, the regularization term includes a first regularization term and a second regularization term; the regularization term determination module includes: a first determination module, used to determine the first regularization term according to the distribution state of the prediction data of the adjacent user groups among the multiple user groups; a second determination module, used to determine the second regularization term according to the prediction data of the target user group among the multiple user groups; the target user group is the user group corresponding to the label with the largest value in the label group.

[0131] In an exemplary embodiment of the present disclosure, the first determination module includes: a difference calculation module, configured to calculate the difference between the prediction data of the adjacent user groups, and determine the first regularization term according to the difference.

[0132] besides, Figure 8 Schematically shows a block diagram of a behavior prediction device according to an exemplary embodiment of the present disclosure. Figure 8 As shown, the behavior prediction device 800 may include:

[0133] The feature information acquisition module 801 is used to acquire information to be processed and acquire multiple feature information of the information to be processed;

[0134] The prediction module 802 is used to input the multiple feature information into the behavior prediction model, predict the user behavior of multiple user groups for the information to be processed, and obtain prediction data corresponding to the multiple user groups; wherein the behavior prediction model is a model trained according to the training method of the behavior prediction model described in any one of claims 1-7.

[0135] It should be noted that the specific details of each module of the training device of the behavior prediction model have been described in detail in the steps of the corresponding behavior prediction model training method, and the specific details of each module of the behavior prediction device have been described in detail in the steps of the corresponding behavior prediction method, so they will not be repeated here.

[0136] Exemplary Electronic Devices

[0137] Refer to the following Fig. 9 The electronic device 900 according to this embodiment of the present disclosure is described. Fig. 9 The electronic device 900 shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0138] like Fig. 9As shown, the electronic device 900 is in the form of a general computing device. The components of the electronic device 900 may include, but are not limited to: the at least one processing unit 910, the at least one storage unit 920, a bus 930 connecting different system components (including the storage unit 920 and the processing unit 109), and a display unit 940. The bus 930 may include a data bus, an address bus, and a control bus.

[0139] The storage unit stores a program code, which can be executed by the processing unit 910, so that the processing unit 910 performs the steps according to various exemplary embodiments of the present disclosure described in the above “Exemplary Method” section of this specification. For example, the processing unit 910 can perform the following steps: Figure 2 or Figure 6 Follow the steps shown in .

[0140] The storage unit 920 may include a readable medium in the form of a volatile storage unit, such as a random access storage unit (RAM) 9201 and / or a cache storage unit 9202 , and may further include a read-only storage unit (ROM) 9203 .

[0141] The storage unit 920 may also include a program / utility 9204 having a set (at least one) of program modules 9205, such program modules 9205 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0142] The electronic device 900 may also communicate with one or more external devices 1000 (e.g., keyboards, pointing devices, Bluetooth devices, etc.). Such communication may be performed via an input / output (I / O) interface 950. Furthermore, the electronic device 900 may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter 960. As shown, the network adapter 960 communicates with other modules of the electronic device 900 via a bus 930. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 900, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0143] Exemplary Media

[0144] Next, a computer-readable storage medium according to an exemplary embodiment of the present disclosure is described.

[0145] In some possible implementations, various aspects of the present disclosure may also be implemented in the form of a program product, which includes a program code. When the program product is run on a terminal device, the program code is used to enable the terminal device to execute the steps of the method for training a behavior prediction model according to various exemplary implementations of the present disclosure described in the above “Exemplary Method” section of this specification. For example, the processing unit may execute the following steps: Figure 2 You can also follow the steps shown in Figure 6 The steps of the behavior prediction method described in .

[0146] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0147] According to the program product for training the behavior prediction model according to the embodiment of the present disclosure, it can adopt a portable compact disk read-only memory (CD-ROM) and include program code, and can be run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited to this. In this document, the readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, an apparatus or a device.

[0148] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, wherein readable program code is carried. Such propagated data signals may take a variety of forms, including, but not limited to, electromagnetic signals, optical signals, or any suitable combination of the foregoing. A readable signal medium may also be any readable medium other than a readable storage medium, which may transmit, propagate, or transfer a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0149] The program code on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0150] Program code for performing the operations of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user computing device, partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0151] It should be noted that although several modules or submodules of the electronic device are mentioned in the above detailed description, this division is only illustrative and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules described above can be embodied in one module. Conversely, the features and functions of one module described above can be further divided into multiple modules to be embodied.

[0152] In addition, although the operations of the disclosed method are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0153] Although the spirit and principle of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined to benefit, and such division is only for the convenience of expression. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A method for training a behavior prediction model, characterized in that: include: Obtaining behavior data of multiple user groups on target information, normalizing the behavior data to obtain normalized behavior data, and generating a tag group for the target information according to the normalized behavior data, wherein the tag group includes tags corresponding to each user group; Inputting the characteristic information of the target information into the behavior prediction model to be trained to obtain a user behavior prediction data group of the target information, wherein the user behavior prediction data group includes user behavior prediction data corresponding to each user group; Determining a loss function of the behavior prediction model according to the tag group, the user behavior prediction data group, and a data distribution state in the user behavior prediction data group; Adjusting the model parameters of the behavior prediction model according to the loss function to train the behavior prediction model; Wherein, determining the loss function of the behavior prediction model according to the tag group, the user behavior prediction data group, and the data distribution state in the user behavior prediction data group includes: Determining an expected loss term according to the tag group and the user behavior prediction data group; Determining a regularization term according to data distribution states of multiple user groups in the user behavior prediction data group; The expected loss term and the regularization term are logically combined to determine the loss function.

2. The method for training a behavior prediction model according to claim 1, characterized in that: The step of normalizing the behavior data of the multiple user groups on the target information to obtain normalized behavior data includes: In the behavior data of each user group for each sample information, determine the maximum value and the minimum value of the global behavior data; The behavior data of each user group on the target information is normalized according to the maximum value of the global behavior data and the minimum value of the global behavior data to obtain the normalized behavior data.

3. The method for training a behavior prediction model according to claim 2, characterized in that: After normalizing the behavior data of each user group on the target information according to the maximum value of the global behavior data and the minimum value of the global behavior data to obtain the normalized behavior data, the method further includes: Determine the local behavior data maximum value and the local behavior data minimum value in the behavior data of each user group for the target information; The normalized behavior data is re-normalized according to the local behavior data maximum value and the local behavior data minimum value to update the normalized behavior data.

4. The method for training a behavior prediction model according to claim 1, characterized in that: The regularization term includes a first regularization term and a second regularization term; The determining of the regularization term according to the data distribution states of the multiple user groups in the user behavior prediction data group includes: determining a first regularization term according to a distribution state of prediction data of adjacent user groups among the multiple user groups; The second regularization term is determined according to the prediction data of a target user group among the multiple user groups; the target user group is the user group corresponding to the label with the largest value in the label group.

5. The method for training a behavior prediction model according to claim 4, characterized in that: The determining the first regularization term according to the distribution state of the prediction data of adjacent user groups among the multiple user groups includes: The difference between the prediction data of the adjacent user groups is calculated, and the first regularization term is determined according to the difference.

6. A behavior prediction method, characterized in that: include: Acquire information to be processed, and acquire multiple feature information of the information to be processed; The multiple feature information is input into a behavior prediction model, and user behaviors of multiple user groups with respect to the information to be processed are predicted to obtain prediction data corresponding to the multiple user groups; wherein the behavior prediction model is a model trained according to the training method of the behavior prediction model according to any one of claims 1-5.

7. A training device for a behavior prediction model, characterized in that: include: A label generation module, used to obtain behavior data of multiple user groups on target information, normalize the behavior data to obtain normalized behavior data, and generate a label group for the target information according to the normalized behavior data, wherein the label group includes labels corresponding to each user group; A prediction data determination module, used for inputting the characteristic information of the target information into the behavior prediction model to be trained to obtain a user behavior prediction data group of the target information, wherein the user behavior prediction data group includes user behavior prediction data corresponding to each user group; A loss function determination module, used to determine the loss function of the behavior prediction model according to the tag group, the user behavior prediction data group and the data distribution state in the user behavior prediction data group; A model training module, used to adjust the model parameters of the behavior prediction model according to the loss function to train the behavior prediction model; The model training module includes: An expected loss term determination module, used to determine an expected loss term according to the tag group and the user behavior prediction data group; A regularization term determination module, used to determine a regularization term according to data distribution states of multiple user groups in the prediction data group; A logical combination module is used to logically combine the expected loss term and the regularization term to determine the loss function.

8. The training device for the behavior prediction model according to claim 7, characterized in that: The normalization module comprises: A global data determination module, used to determine the maximum value and the minimum value of global behavior data in the behavior data of each user group for each sample information; The normalization control module is used to normalize the behavior data of each user group on the target information according to the maximum value and the minimum value of the global behavior data to obtain the normalized behavior data.

9. The training device for the behavior prediction model according to claim 8, characterized in that: After normalizing the behavior data of each user group on the target information according to the maximum value of the global behavior data and the minimum value of the global behavior data to obtain the normalized behavior data, the device further includes: A local data determination module, used to determine the maximum value and the minimum value of local behavior data in the behavior data of each user group on the target information; The normalization update module is used to perform re-normalization processing on the normalized behavior data according to the maximum value of the local behavior data and the minimum value of the local behavior data to update the normalized behavior data.

10. The behavior prediction model training device according to claim 9, characterized in that: The regularization term includes a first regularization term and a second regularization term; the regularization term determination module includes: A first determination module, configured to determine a first regularization term according to a distribution state of prediction data of adjacent user groups among the multiple user groups; The second determination module is used to determine the second regularization term according to the prediction data of the target user group among the multiple user groups; the target user group is the user group corresponding to the label with the largest value in the label group.

11. The training device for the behavior prediction model according to claim 10, characterized in that: The first determining module comprises: The difference calculation module is used to calculate the difference between the prediction data of the adjacent user groups and determine the first regularization term according to the difference.

12. A behavior prediction device, characterized in that: include: A feature information acquisition module, used to acquire information to be processed and to acquire a plurality of feature information of the information to be processed; A prediction module is used to input the multiple feature information into a behavior prediction model, predict the user behavior of multiple user groups for the information to be processed, and obtain prediction data corresponding to the multiple user groups; wherein the behavior prediction model is a model trained according to the training method of the behavior prediction model according to any one of claims 1-5.

13. An electronic device, characterized in that: include: processor; as well as A memory for storing executable instructions; Wherein, the processor is configured to execute the behavior prediction model training method described in any one of claims 1 to 5 or the behavior prediction method described in claim 6 by executing the executable instructions.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the training method of the behavior prediction model as described in any one of claims 1 to 5 or the behavior prediction method as described in claim 6.

Images