Content recommendation model training method and device, content recommendation method and device and electronic equipment

Abstract

The invention relates to a content recommendation model training method and device, a content recommendation method and device, electronic equipment, a storage medium and a computer program product. The method comprises the following steps: acquiring a sample cross cluster and a sample data set of the sample cross cluster; inputting the sample data set of each sample cross cluster into a to-be-trained content recommendation model to obtain an interaction index prediction score of the sample data set of each sample cross cluster; according to the interaction index prediction score of the sample data set of each sample cross cluster, determining a target loss value of each sample cross cluster, and performing fusion processing on the target loss values of each sample cross cluster to obtain a first loss value; determining a second loss value according to the first interaction index prediction score of the first sample data of each sample cross cluster; and training a to-be-trained content recommendation model according to the first loss value and the second loss value to obtain a trained content recommendation model. By adopting the method, the content recommendation accuracy can be improved.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a content recommendation model training method, a content recommendation method, an apparatus, an electronic device, a storage medium, and a computer program product. Background Art

[0002] With the development of computer technology, many recommendation models have emerged. The recommendation system will recommend content to accounts through recommendation models, such as pictures, live broadcasts, videos, etc.

[0003] In the related art, when training a recommendation model, modeling is generally performed for account clusters or content clusters to achieve account fairness or content fairness; however, considering account clusters or content clusters alone can easily miss potential interests in the system, resulting in low accuracy of the output results of the trained recommendation model, and thus low accuracy of content recommendations. Summary of the invention

[0004] The present disclosure provides a content recommendation model training method, a content recommendation method, an apparatus, an electronic device, a storage medium, and a computer program product to at least solve the problem of low content recommendation accuracy in the related art. The technical solution of the present disclosure is as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a content recommendation model training method is provided, comprising:

[0006] Obtain sample cross clusters and sample data groups of the sample cross clusters; each sample cross cluster includes a sample account cluster and a sample content cluster; each sample data group includes first sample data and second sample data; the first sample data includes a sample account in the sample account cluster in the sample cross cluster to which it belongs, and a sample content with which the sample account has interacted in the sample content cluster in the sample cross cluster to which it belongs; the second sample data includes the sample account and a sample content with which the sample account has not interacted in each of the sample cross clusters;

[0007] Inputting the sample data group of each sample cross cluster into the content recommendation model to be trained, and obtaining the interaction index prediction score corresponding to the sample data group of each sample cross cluster; the interaction index prediction score includes the first interaction index prediction score corresponding to the first sample data, and the second interaction index prediction score corresponding to the second sample data;

[0008] Determine the target loss value corresponding to each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster, and perform fusion processing on the target loss value corresponding to each sample cross cluster to obtain a first loss value;

[0009] Determining a second loss value according to a first interaction indicator prediction score corresponding to the first sample data of each sample cross cluster;

[0010] The content recommendation model to be trained is iteratively trained according to the first loss value and the second loss value to obtain a trained content recommendation model.

[0011] In an exemplary embodiment, determining the target loss value corresponding to each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster includes:

[0012] Determine the initial loss value corresponding to each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster;

[0013] According to the initial loss value corresponding to each sample cross cluster, the model parameters of the content recommendation model to be trained are updated to obtain an updated content recommendation model corresponding to each sample cross cluster;

[0014] Outputting the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster through the updated content recommendation model corresponding to each sample cross cluster;

[0015] According to the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster, the updated loss value corresponding to the sample data group of each sample cross cluster is determined, and the updated loss values ​​corresponding to the sample data group of each sample cross cluster are fused respectively to obtain the target loss value corresponding to each sample cross cluster.

[0016] In an exemplary embodiment, determining the initial loss value corresponding to each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster includes:

[0017] Determine the loss value corresponding to the sample data group of each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster;

[0018] The loss values ​​corresponding to the sample data groups of each sample cross cluster are fused respectively to obtain the initial loss value corresponding to each sample cross cluster.

[0019] In an exemplary embodiment, updating the model parameters of the content recommendation model to be trained according to the initial loss value corresponding to each sample cross cluster to obtain the updated content recommendation model corresponding to each sample cross cluster includes:

[0020] Determine the initial model parameter update gradient corresponding to each sample cross cluster according to the initial loss value corresponding to each sample cross cluster;

[0021] According to the current model parameters of the content recommendation model to be trained, the norm of the current model parameters and the initial model parameters corresponding to each sample cross cluster, the gradient is updated to obtain the updated model parameters corresponding to each sample cross cluster;

[0022] According to the updated model parameters corresponding to each sample cross cluster, the current model parameters of the content recommendation model to be trained are updated to obtain the updated content recommendation model corresponding to each sample cross cluster.

[0023] In an exemplary embodiment, the step of fusing the target loss value corresponding to each sample cross cluster to obtain a first loss value includes:

[0024] Determine the current weight corresponding to each sample cross cluster;

[0025] According to the current weight corresponding to each sample cross cluster, the target loss value corresponding to each sample cross cluster is fused to obtain a first loss value.

[0026] In an exemplary embodiment, determining the current weight corresponding to each sample cross cluster includes:

[0027] Determine, according to the target loss value corresponding to each sample cross cluster, the target model parameter update gradient corresponding to each sample cross cluster;

[0028] Determine the cumulative contribution information corresponding to each sample cross cluster according to the target loss value and the target model parameter update gradient corresponding to each sample cross cluster; the cumulative contribution information corresponding to each sample cross cluster is used to characterize the fusion processing result of the difference value between the updated target loss value corresponding to each sample cross cluster and the target loss value when the target model parameter update gradient of each sample cross cluster is used to update the content recommendation model to be trained;

[0029] According to the accumulated contribution information corresponding to each sample cross cluster, the historical weight corresponding to each sample cross cluster is updated to obtain the current weight corresponding to each sample cross cluster.

[0030] In an exemplary embodiment, determining the cumulative contribution information corresponding to each sample cross cluster according to the target loss value and the target model parameter update gradient corresponding to each sample cross cluster includes:

[0031] Determine the contribution information of each sample cross cluster to each of the sample cross clusters according to the target model parameter update gradient corresponding to each sample cross cluster; the contribution information of each sample cross cluster to each of the sample cross clusters is used to characterize the difference between the updated target loss value and the target loss value corresponding to each of the sample cross clusters when the content recommendation model to be trained is updated using the target model parameter update gradient of each sample cross cluster;

[0032] Determining, according to the target loss value corresponding to each sample cross cluster, the weight of the contribution information of each sample cross cluster to each of the sample cross clusters;

[0033] According to the weight of the contribution information of each sample cross cluster to each sample cross cluster, the contribution information of each sample cross cluster to each sample cross cluster is fused to obtain the cumulative contribution information corresponding to each sample cross cluster.

[0034] In an exemplary embodiment, determining the second loss value according to the first interaction indicator prediction score corresponding to the first sample data of each sample cross cluster includes:

[0035] Determining an average interaction index prediction score of each sample cross cluster according to the first interaction index prediction score corresponding to the first sample data of each sample cross cluster;

[0036] The second loss value is determined according to the average interaction index prediction score of each sample cross cluster.

[0037] According to a second aspect of an embodiment of the present disclosure, a content recommendation method is provided, including:

[0038] Obtain the recommended content corresponding to the account to be recommended;

[0039] Inputting the account features of the to-be-recommended account and the content features of the to-be-recommended content into a trained content recommendation model for prediction processing to obtain a predicted score of the interaction index of the to-be-recommended content; the trained content recommendation model is trained by the content recommendation model training method described in any one of the embodiments of the first aspect;

[0040] Filter out target content from the content to be recommended according to the interaction indicator prediction score, and recommend the target content to the account to be recommended.

[0041] According to a third aspect of an embodiment of the present disclosure, a content recommendation model training device is provided, including:

[0042] A data acquisition unit is configured to acquire sample cross clusters and sample data groups of the sample cross clusters; each sample cross cluster includes a sample account cluster and a sample content cluster; each sample data group includes first sample data and second sample data; the first sample data includes a sample account in the sample account cluster in the sample cross cluster to which it belongs, and a sample content with which the sample account has interacted in the sample content cluster in the sample cross cluster to which it belongs; the second sample data includes the sample account and a sample content with which the sample account has not interacted in each of the sample cross clusters;

[0043] A score determination unit is configured to input the sample data group of each sample cross cluster into the content recommendation model to be trained, and obtain the interaction index prediction score corresponding to the sample data group of each sample cross cluster; the interaction index prediction score includes a first interaction index prediction score corresponding to the first sample data, and a second interaction index prediction score corresponding to the second sample data;

[0044] A first calculation unit is configured to perform an interaction index prediction score corresponding to the sample data group of each sample cross cluster, determine a target loss value corresponding to each sample cross cluster, and perform a fusion process on the target loss value corresponding to each sample cross cluster to obtain a first loss value;

[0045] A second calculation unit is configured to determine a second loss value according to a first interaction indicator prediction score corresponding to the first sample data of each sample cross cluster;

[0046] The model training unit is configured to perform iterative training on the content recommendation model to be trained according to the first loss value and the second loss value to obtain a trained content recommendation model.

[0047] In an exemplary embodiment, the first computing unit is further configured to determine the initial loss value corresponding to each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster; update the model parameters of the content recommendation model to be trained according to the initial loss value corresponding to each sample cross cluster, and obtain the updated content recommendation model corresponding to each sample cross cluster; output the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster through the updated content recommendation model corresponding to each sample cross cluster; determine the updated loss value corresponding to the sample data group of each sample cross cluster according to the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster, and fuse the updated loss values ​​corresponding to the sample data group of each sample cross cluster to obtain the target loss value corresponding to each sample cross cluster.

[0048] In an exemplary embodiment, the first computing unit is further configured to determine the loss value corresponding to the sample data group of each sample cross cluster based on the interaction index prediction score corresponding to the sample data group of each sample cross cluster; and to perform fusion processing on the loss values ​​corresponding to the sample data group of each sample cross cluster respectively to obtain the initial loss value corresponding to each sample cross cluster.

[0049] In an exemplary embodiment, the first computing unit is further configured to determine the initial model parameter update gradient corresponding to each sample cross cluster according to the initial loss value corresponding to each sample cross cluster; obtain the updated model parameters corresponding to each sample cross cluster according to the current model parameters of the content recommendation model to be trained, the norm of the current model parameters and the initial model parameter update gradient corresponding to each sample cross cluster; and update the current model parameters of the content recommendation model to be trained according to the updated model parameters corresponding to each sample cross cluster, to obtain the updated content recommendation model corresponding to each sample cross cluster.

[0050] In an exemplary embodiment, the first computing unit is further configured to determine the current weight corresponding to each sample cross cluster; and according to the current weight corresponding to each sample cross cluster, perform fusion processing on the target loss value corresponding to each sample cross cluster to obtain a first loss value.

[0051] In an exemplary embodiment, the first computing unit is further configured to determine the target model parameter update gradient corresponding to each sample cross cluster according to the target loss value corresponding to each sample cross cluster; determine the cumulative contribution information corresponding to each sample cross cluster according to the target loss value and the target model parameter update gradient corresponding to each sample cross cluster; the cumulative contribution information corresponding to each sample cross cluster is used to characterize the fusion processing result of the difference value between the updated target loss value and the target loss value corresponding to each sample cross cluster when the target model parameter update gradient of each sample cross cluster is used to update the content recommendation model to be trained; according to the cumulative contribution information corresponding to each sample cross cluster, the historical weight corresponding to each sample cross cluster is updated to obtain the current weight corresponding to each sample cross cluster.

[0052] In an exemplary embodiment, the first computing unit is further configured to execute, according to the target model parameter update gradient corresponding to each sample cross cluster, to determine the contribution information of each sample cross cluster to each of the sample cross clusters; the contribution information of each sample cross cluster to each of the sample cross clusters is used to characterize the difference between the updated target loss value and the target loss value corresponding to each of the sample cross clusters when the content recommendation model to be trained is updated using the target model parameter update gradient of each sample cross cluster; according to the target loss value corresponding to each sample cross cluster, to determine the weight of the contribution information of each sample cross cluster to each of the sample cross clusters; according to the weight of the contribution information of each sample cross cluster to each of the sample cross clusters, to fuse the contribution information of each sample cross cluster to each of the sample cross clusters to obtain the cumulative contribution information corresponding to each sample cross cluster.

[0053] In an exemplary embodiment, the second computing unit is further configured to determine the average interaction index prediction score of each sample cross-cluster based on the first interaction index prediction score corresponding to the first sample data of each sample cross-cluster; and determine the second loss value based on the average interaction index prediction score of each sample cross-cluster.

[0054] According to a fourth aspect of an embodiment of the present disclosure, there is provided a content recommendation device, including:

[0055] A content acquisition unit is configured to acquire the content to be recommended corresponding to the account to be recommended;

[0056] a score output unit configured to input the account features of the to-be-recommended account and the content features of the to-be-recommended content into a trained content recommendation model for prediction processing to obtain a predicted score of the interaction index of the to-be-recommended content; the trained content recommendation model is trained by the content recommendation model training method described in any one of the embodiments of the first aspect;

[0057] The content recommendation unit is configured to filter out target content from the content to be recommended according to the interaction indicator prediction score, and recommend the target content to the account to be recommended.

[0058] According to a fifth aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0059] processor;

[0060] a memory for storing instructions executable by the processor;

[0061] The processor is configured to execute the instructions to implement the content recommendation model training method as described in any one of the embodiments of the first aspect, or the content recommendation method as described in any one of the embodiments of the second aspect.

[0062] According to a sixth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the content recommendation model training method described in any one of the embodiments of the first aspect, or the content recommendation method described in any one of the embodiments of the second aspect.

[0063] According to a seventh aspect of the embodiments of the present disclosure, a computer program product is provided, wherein the computer program product includes instructions, and when the instructions are executed by a processor of an electronic device, the electronic device is capable of executing the content recommendation model training method described in any one of the embodiments of the first aspect, or the content recommendation method described in any one of the embodiments of the second aspect.

[0064] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0065] Obtain sample cross-cluster and sample data groups of sample cross-cluster; each sample cross-cluster includes a sample account cluster and a sample content cluster; each sample data group includes first sample data and second sample data; the first sample data includes a sample account in the sample account cluster in the sample cross-cluster to which it belongs, and a sample content with which the sample account has interacted in the sample content cluster in the sample cross-cluster to which it belongs; the second sample data includes the sample account and a sample content with which the sample account has not interacted in each sample cross-cluster; then input the sample data group of each sample cross-cluster into the content recommendation model to be trained, and obtain the sample data group pair of each sample cross-cluster The corresponding interaction indicator prediction score; the interaction indicator prediction score includes the first interaction indicator prediction score corresponding to the first sample data, and the second interaction indicator prediction score corresponding to the second sample data; and according to the interaction indicator prediction score corresponding to the sample data group of each sample cross-cluster, the target loss value corresponding to each sample cross-cluster is determined, and the target loss value corresponding to each sample cross-cluster is fused to obtain the first loss value, and according to the first interaction indicator prediction score corresponding to the first sample data of each sample cross-cluster, the second loss value is determined; finally, according to the first loss value and the second loss value, the content recommendation model to be trained is iteratively trained to obtain the trained content recommendation model. In this way, in the process of training the content recommendation model, modeling is performed for sample cross-clusters, and the first loss value constructed based on the interaction index prediction score corresponding to the sample data group of each sample cross-cluster is comprehensively considered, so that the trained content recommendation model can have similar discrimination ability for positive and negative examples of different cross-clusters; at the same time, the second loss value constructed based on the first interaction index prediction score corresponding to the first sample data of each sample cross-cluster is considered, so that the trained content recommendation model can align the interaction index prediction scores of positive examples of different cross-clusters to prevent the positive examples of a certain cross-cluster from being systematically ranked behind the positive examples of another cross-cluster, which is conducive to exploring the potential interests of cross-clusters, thereby making the output results of the trained content recommendation model more accurate, thereby improving the accuracy of content recommendation.

[0066] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0068] Figure 1 The figure is a schematic diagram showing a cross-bilateral unfairness phenomenon in a recommendation system according to an exemplary embodiment.

[0069] Figure 2 The figure is a flowchart of a content recommendation model training method according to an exemplary embodiment.

[0070] Figure 3 is a flowchart showing the steps of determining a target loss value corresponding to each sample cross cluster according to an exemplary embodiment.

[0071] Figure 4 The present invention is a flowchart of a content recommendation method according to an exemplary embodiment.

[0072] Figure 5 The figure is a flowchart of another content recommendation method according to an exemplary embodiment.

[0073] Figure 6 is a target schematic diagram of a training method for cross-bilateral fair recommendation according to an exemplary embodiment.

[0074] Figure 7 The present invention is a block diagram of a content recommendation model training device according to an exemplary embodiment.

[0075] Figure 8 The present invention is a block diagram of a content recommendation device according to an exemplary embodiment.

[0076] Fig. 9 It is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0077] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0078] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0079] It should also be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data for analysis, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0080] Since recommendation systems involve resource allocation, the fairness of recommendations has received increasing attention in recent years. Unfair recommendation systems often lead to the inability to accurately model some user preferences and high-quality content, further leading to a decline in user experience, loss of creators, and loss of diversity in the system ecosystem. Existing recommendation fairness methods only consider separate modeling of user fairness and content fairness, ignoring the fairness of potential user-content cross-groups. For example, refer to Figure 1 , (1) shows the distribution of user content interests in the system; (2) examines a simple recommendation strategy: recommending horror movies only to male users and romance movies only to female users; (3) from the perspective of existing bilateral fairness methods, it can be found that both the user side and the content side are fair. From the user side, the satisfaction rate of recommendation interests for male and female users is 90%, while from the content side, 90% of the recommendations for horror movies and romance movies are effective (i.e., they are recommended to satisfied audiences); (4) from the perspective of cross-bilateral groups, it can be found that the recommendation method completely ignores the potential interests in the system, i.e., male-romance movies and female-horror movies, and there is a problem of cross-bilateral unfairness. This cross-bilateral unfairness affects user satisfaction, hinders the distribution of content to potential audiences, and is not conducive to the development of a diverse platform ecosystem and the discovery of potential markets. It can be seen that traditional methods are designed for the entire user group and the entire content group, and do not take into account the cross-bilateral groups of users and content. As Figure 1 As shown, this method targeting the entire user group or content group cannot effectively identify and alleviate potential cross-bilateral unfairness problems. Based on this, the present disclosure proposes a content recommendation model training method to solve this problem.

[0081] Figure 2 is a flowchart of a content recommendation model training method according to an exemplary embodiment. Figure 2 As shown, the content recommendation model training method is used in a server; it is understandable that the method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The terminal can be, but is not limited to, various personal computers, laptops, smart phones, and tablet computers, and the server can be implemented as an independent server or a server cluster composed of multiple servers. In this exemplary embodiment, the method includes the following steps:

[0082] In step S210 , sample cross clusters and sample data groups of the sample cross clusters are obtained.

[0083] Among them, each sample cross-cluster includes a sample account cluster and a sample content cluster; each sample data group includes first sample data and second sample data; the first sample data includes a sample account in the sample account cluster in the sample cross-cluster to which it belongs, and a sample content with which the sample account has interacted in the sample content cluster in the sample cross-cluster to which it belongs; the second sample data includes the sample account and a sample content with which the sample account has not interacted in each sample cross-cluster.

[0084] Among them, the sample cross cluster refers to a cross bilateral group composed of a sample account cluster and a sample content cluster. In actual scenarios, by clustering sample accounts and sample contents, multiple sample account clusters and multiple sample content clusters can be obtained, and then multiple sample account clusters and multiple sample content clusters can be cross-combined to obtain multiple sample cross clusters; for example, the sample account clusters are A1 and A2, and the sample content clusters are B1 and B2, then the sample cross clusters formed are A1B1, A1B2, A2B1 and A2B2.

[0085] Among them, the sample account cluster refers to a group of sample accounts with similar account characteristics, specifically a group of sample accounts with similar account attributes, such as the same gender, similar occupation, similar hobbies, etc. For example, sample account cluster A1 includes sample accounts U1 and U2, and sample account cluster A2 includes sample accounts U3 and U4. In actual scenarios, sample accounts refer to application accounts, such as short video accounts, browser accounts, etc.

[0086] The sample content cluster refers to a group of sample contents with similar content features, specifically a group of sample contents with similar content attributes, such as similar content types and similar content themes. For example, sample content cluster B1 includes sample contents I1 and I2, and sample content cluster B2 includes sample contents I3 and I4. In actual scenarios, sample content refers to content on the Internet, such as videos, articles, advertisements, pictures, live broadcasts, etc.

[0087] Among them, each sample cross cluster includes multiple sample data groups, and each sample data group includes a first sample data and a second sample data. For a certain sample cross cluster, the first sample data refers to a positive example of account-content interaction, specifically including a sample account in the sample account cluster in the sample cross cluster, and a sample content that the sample account has interacted with (such as clicked, liked, collected, etc.) in the sample content cluster in the sample cross cluster; the second sample data refers to a negative example of account-content interaction, specifically including the sample account, and any sample content that the sample account has not interacted with in all cross clusters. For example, for the sample cross cluster A1B1, the sample account U1 has interacted with sample content I1 and I3, but not with I2 and I4, and the sample account U2 has interacted with sample content I2 and I4, but not with I1 and I3, then there are two sample data groups corresponding to the sample cross cluster A1B1, namely: [(U1, I1), (U1, I4)], [(U2, I2), (U2, I3)]. In this way, when constructing the sample data group, preference learning can be performed by distinguishing between content that the sample account has interacted with and content that it has not interacted with. It should be noted that the second sample data in the first sample data group can be (U1, I4) or (U1, I2), and the second sample data in the second sample data group can be (U2, I3) or (U2, I1); in this way, for the content in the account-content interaction negative example, all sample content clusters (such as B1 and B2) can be covered, not limited to the current sample content cluster (such as B1).

[0088] Exemplarily, the server obtains multiple sample accounts, multiple sample contents, and interaction records of the sample accounts with the sample contents; clusters the multiple sample accounts to obtain multiple sample account clusters; clusters the multiple sample contents to obtain multiple sample content clusters; cross-combines the multiple sample account clusters and the multiple sample content clusters to obtain multiple sample cross-clustering clusters; determines multiple groups of first sample data and second sample data corresponding to each sample cross-cluster based on the interaction records of the sample accounts with the sample contents, and constructs multiple sample data groups corresponding to each sample cross-cluster based on the multiple groups of first sample data and second sample data corresponding to each sample cross-cluster.

[0089] In step S220, the sample data group of each sample cross-cluster is input into the content recommendation model to be trained to obtain the interaction indicator prediction score corresponding to the sample data group of each sample cross-cluster; the interaction indicator prediction score includes the first interaction indicator prediction score corresponding to the first sample data, and the second interaction indicator prediction score corresponding to the second sample data.

[0090] The content recommendation model to be trained may be a neural network model, a deep learning model, a large model, etc. The interaction index is the interaction index of the sample account to the sample content, such as clicks, likes, collections, etc. The interaction index prediction score refers to the normalized interaction index prediction score, such as click prediction score, like prediction score, collection prediction score, etc.

[0091] Among them, the first interaction indicator prediction score corresponding to the first sample data refers to the normalized interaction indicator prediction score corresponding to the account-content interaction positive example; the second interaction indicator prediction score corresponding to the second sample data refers to the normalized interaction indicator prediction score corresponding to the account-content interaction negative example.

[0092] Exemplarily, the server obtains the account information of the sample accounts in the sample data group of each sample cross-cluster, the content information of the interacted sample content, and the content information of the non-interacted sample content, and performs feature extraction processing on the account information of the sample accounts, the content information of the interacted sample content, and the content information of the non-interacted sample content to obtain sample account features, interacted sample content features, and non-interacted sample content features; then, the server inputs the sample account features and the interacted sample content features into the content recommendation model to be trained, performs a series of processing on the sample account features and the interacted sample content features through the content recommendation model to be trained, obtains the interaction index prediction score of the sample account for the interacted sample content, and performs normalization processing on the sample account to obtain the normalized interaction index prediction score of the sample account for the interacted sample content, as the first a first interaction indicator prediction score corresponding to one sample data; similarly, the server inputs the sample account characteristics and the non-interaction sample content characteristics into the content recommendation model to be trained, and performs a series of processing on the sample account characteristics and the non-interaction sample content characteristics through the content recommendation model to be trained to obtain the interaction indicator prediction score of the sample account for the non-interaction sample content, and performs normalization processing on it to obtain the normalized interaction indicator prediction score of the sample account for the non-interaction sample content as the second interaction indicator prediction score corresponding to the second sample data; according to the first interaction indicator prediction score corresponding to the first sample data and the second interaction indicator prediction score corresponding to the second sample data, the interaction indicator prediction score corresponding to the sample data group is obtained; and so on, the interaction indicator prediction score corresponding to the sample data group of each sample cross cluster can be obtained.

[0093] In step S230, the target loss value corresponding to each sample cross cluster is determined according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster, and the target loss value corresponding to each sample cross cluster is fused to obtain a first loss value.

[0094] The target loss value refers to the sharpness perception loss, specifically the worst loss within a certain range of the current model parameters, such as the worst BPR (Bayesian Personalized Ranking) loss within a certain range of the current model parameters. Each sample cross cluster corresponds to a target loss value, which is calculated based on the interaction indicator prediction score corresponding to the sample data group of each sample cross cluster.

[0095] The first loss value refers to the collaborative loss of sharpness perception, specifically the weighted sum of the target loss values ​​corresponding to each sample cross cluster. By training the content recommendation model to be trained with the first loss value, the trained content recommendation model can have similar ability to distinguish positive and negative examples of different cross clusters.

[0096] Exemplarily, the server calculates the sharpness perception loss corresponding to each sample cross cluster based on the interaction indicator prediction score corresponding to the sample data group of each sample cross cluster and the sharpness perception loss function, as the target loss value corresponding to each sample cross cluster; and performs weighted summation processing on the target loss value corresponding to each sample cross cluster to obtain a first loss value.

[0097] In step S240, a second loss value is determined according to the first interaction indicator prediction score corresponding to the first sample data of each sample cross cluster.

[0098] The second loss value refers to the subgroup alignment loss, which is calculated based on the first interaction index prediction score corresponding to the first sample data of each sample cross cluster. By training the content recommendation model to be trained with the second loss value, the trained content recommendation model can align the interaction index prediction scores of positive examples of different cross clusters to prevent the positive examples of a certain cross cluster from being systematically ranked behind the positive examples of another cross cluster.

[0099] Exemplarily, the server calculates the average interaction index prediction score of each sample cross cluster based on the first interaction index prediction score corresponding to the first sample data of each sample cross cluster, and determines the second loss value based on the average interaction index prediction score of each sample cross cluster.

[0100] In step S250, the content recommendation model to be trained is iteratively trained according to the first loss value and the second loss value to obtain a trained content recommendation model.

[0101] Exemplarily, the server fuses the first loss value and the second loss value to obtain a total loss value; adjusts the model parameters of the content recommendation model to be trained according to the total loss value, and continues to repeat steps S220 to S250 until the training end condition is met, such as the current number of training times reaches a preset number of training times, the total loss value obtained based on the trained content recommendation model is less than a preset threshold, etc.; finally, the trained content recommendation model that meets the training end condition is used as the content recommendation model that has completed training.

[0102] For example, the total loss value of the content recommendation model to be trained is calculated by the following formula 1:

[0103]

[0104] in, represents the total loss value, represents the first loss value, λ represents the hyperparameter, Represents the second loss value.

[0105] In the above-mentioned content recommendation model training method, sample cross-cluster and sample data groups of sample cross-cluster are obtained; each sample cross-cluster includes a sample account cluster and a sample content cluster; each sample data group includes first sample data and second sample data; the first sample data includes a sample account in the sample account cluster in the sample cross-cluster to which it belongs, and a sample content with which the sample account has interacted in the sample content cluster in the sample cross-cluster to which it belongs; the second sample data includes the sample account and a sample content with which the sample account has not interacted in each sample cross-cluster; then the sample data group of each sample cross-cluster is input into the content recommendation model to be trained, and each sample cross-cluster is obtained. The interaction indicator prediction score corresponding to the sample data group includes a first interaction indicator prediction score corresponding to the first sample data and a second interaction indicator prediction score corresponding to the second sample data; and according to the interaction indicator prediction score corresponding to the sample data group of each sample cross-cluster, the target loss value corresponding to each sample cross-cluster is determined, and the target loss value corresponding to each sample cross-cluster is fused to obtain a first loss value, and according to the first interaction indicator prediction score corresponding to the first sample data of each sample cross-cluster, the second loss value is determined; finally, according to the first loss value and the second loss value, the content recommendation model to be trained is iteratively trained to obtain the trained content recommendation model. In this way, in the process of training the content recommendation model, modeling is performed for sample cross-clusters, and the first loss value constructed based on the interaction index prediction score corresponding to the sample data group of each sample cross-cluster is comprehensively considered, so that the trained content recommendation model can have similar discrimination ability for positive and negative examples of different cross-clusters; at the same time, the second loss value constructed based on the first interaction index prediction score corresponding to the first sample data of each sample cross-cluster is considered, so that the trained content recommendation model can align the interaction index prediction scores of positive examples of different cross-clusters to prevent the positive examples of a certain cross-cluster from being systematically ranked behind the positive examples of another cross-cluster, which is conducive to exploring the potential interests of cross-clusters, thereby making the output results of the trained content recommendation model more accurate, thereby improving the accuracy of content recommendation.

[0106] In an exemplary embodiment, if Figure 3 As shown, the above step S230 determines the target loss value corresponding to each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster, which can be specifically achieved by the following steps:

[0107] In step S310, an initial loss value corresponding to each sample cross cluster is determined according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster.

[0108] In step S320, the model parameters of the content recommendation model to be trained are updated according to the initial loss value corresponding to each sample cross cluster, so as to obtain an updated content recommendation model corresponding to each sample cross cluster.

[0109] In step S330, the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster is outputted through the updated content recommendation model corresponding to each sample cross cluster.

[0110] In step S340, based on the update interaction index prediction score corresponding to the sample data group of each sample cross cluster, the update loss value corresponding to the sample data group of each sample cross cluster is determined, and the update loss value corresponding to the sample data group of each sample cross cluster is fused respectively to obtain the target loss value corresponding to each sample cross cluster.

[0111] The initial loss value refers to BPR (Bayesian Personalized Ranking) loss, cross entropy loss, Hinge loss, etc. Each sample cross cluster corresponds to an initial loss value, which is calculated based on the interaction index prediction score corresponding to the sample data group of each sample cross cluster.

[0112] Among them, each sample cross cluster corresponds to an updated content recommendation model, and the updated content recommendation model corresponding to each sample cross cluster is obtained by updating the content recommendation model to be trained according to the initial loss value corresponding to each sample cross cluster.

[0113] The updated interaction index prediction score corresponding to the sample data group of each sample cross cluster refers to the interaction index prediction score corresponding to the sample data group of each sample cross cluster output by the updated content recommendation model corresponding to each sample cross cluster.

[0114] Among them, the target loss value corresponding to each sample cross cluster is calculated according to the update interaction index prediction score corresponding to the sample data group of each sample cross cluster.

[0115] Exemplarily, the server calculates the BPR loss corresponding to each sample cross cluster based on the interaction index prediction score corresponding to the sample data group of each sample cross cluster, as the initial loss value corresponding to each sample cross cluster, for example, sample cross clusters A1B1, A1B2, A2B1, and A2B2 each correspond to an initial loss value. Next, the server determines the updated model parameters corresponding to each sample cross cluster based on the initial loss value corresponding to each sample cross cluster, and updates the model parameters of the content recommendation model to be trained based on the updated model parameters corresponding to each sample cross cluster, to obtain an updated content recommendation model corresponding to each sample cross cluster; for example, sample cross clusters A1B1, A1B2, A2B1, and A2B2 each correspond to an updated content recommendation model. Then, the server inputs the sample data group of each sample cross cluster into the updated content recommendation model corresponding to each sample cross cluster, and performs a series of processing on the sample data group of each sample cross cluster through the updated content recommendation model corresponding to each sample cross cluster, so as to obtain the update interaction index prediction score corresponding to the sample data group of each sample cross cluster; for example, the sample data group of sample cross cluster A1B1 is input into the updated content recommendation model corresponding to A1B1, so as to obtain the update interaction index prediction score corresponding to the sample data group of sample cross cluster A1B1; the sample data group of sample cross cluster A1B2 is input into the updated content recommendation model corresponding to A1B2, so as to obtain the update interaction index prediction score corresponding to the sample data group of sample cross cluster A1B2, etc. Finally, the server determines the update loss value (such as update BPR loss) corresponding to the sample data group of each sample cross cluster according to the update interaction index prediction score corresponding to the sample data group of each sample cross cluster; the update loss value corresponding to the sample data group of each sample cross cluster is fused (such as averaging), so as to obtain the update loss value (such as update BPR loss) corresponding to each sample cross cluster as the target loss value corresponding to each sample cross cluster. For example, for the sample cross cluster A1B1, the update loss value corresponding to the sample data group [(U1, I1), (U1, I4)] and the update loss value corresponding to the sample data group [(U2, I2), (U2, I3)] are added and averaged to obtain the target loss value corresponding to the sample cross cluster A1B1.

[0116] The technical solution provided by the embodiment of the present disclosure first determines the initial loss value corresponding to each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster, then updates the model parameters of the content recommendation model to be trained according to the initial loss value corresponding to each sample cross cluster, and obtains the updated content recommendation model corresponding to each sample cross cluster, and then outputs the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster through the updated content recommendation model corresponding to each sample cross cluster, and finally determines the target loss value corresponding to each sample cross cluster according to the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster; in this way, after the content recommendation model is updated, the target loss value corresponding to each sample cross cluster is calculated, which is conducive to improving the accuracy of determining the target loss value corresponding to each sample cross cluster; at the same time, in the model training process, the target loss value corresponding to each sample cross cluster is comprehensively considered, which is conducive to improving the content recommendation accuracy of the content recommendation model.

[0117] In an exemplary embodiment, the above-mentioned step S310 determines the initial loss value corresponding to each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster, and specifically includes the following contents: determining the loss value corresponding to the sample data group of each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster; fusing the loss values ​​corresponding to the sample data group of each sample cross cluster respectively to obtain the initial loss value corresponding to each sample cross cluster.

[0118] Among them, each sample data group corresponds to a loss value, such as the triple BPR loss, which is calculated based on the interaction indicator prediction score corresponding to the sample data group (i.e., the first interaction indicator prediction score corresponding to the first sample data and the second interaction indicator prediction score corresponding to the second sample data).

[0119] The fusing process of the loss values ​​corresponding to the sample data group of each sample cross cluster refers to averaging these loss values.

[0120] Exemplarily, the server determines the loss value (such as BPR loss) corresponding to the sample data group of each sample cross cluster based on the interaction indicator prediction score corresponding to the sample data group of each sample cross cluster; the loss values ​​corresponding to the sample data group of each sample cross cluster are respectively fused (such as averaging) to obtain the loss value (such as BPR loss) corresponding to each sample cross cluster as the initial loss value corresponding to each sample cross cluster.

[0121] For example, for the sample cross cluster A1B1, the server calculates the loss value S1 corresponding to the sample data group [(U1,I1),(U1,I4)] according to the interaction indicator prediction score corresponding to the sample data group [(U1,I1),(U1,I4)]; calculates the loss value S2 corresponding to the sample data group [(U2,I2),(U2,I3)] according to the interaction indicator prediction score corresponding to the sample data group [(U2,I2),(U2,I3)]; then, the loss value S1 corresponding to the sample data group [(U1,I1),(U1,I4)] and the loss value S2 corresponding to the sample data group [(U2,I2),(U2,I3)] are added and averaged to obtain the initial loss value corresponding to the sample cross cluster A1B1, that is, (S1+S2) / 2.

[0122] The technical solution provided by the embodiment of the present disclosure first determines the loss value corresponding to the sample data group of each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster, and then fuses the loss values ​​corresponding to the sample data group of each sample cross cluster respectively to obtain the initial loss value corresponding to each sample cross cluster; in this way, when calculating the initial loss value corresponding to each sample cross cluster, the loss value corresponding to the sample data group of each sample cross cluster is comprehensively considered, which is conducive to improving the accuracy of determining the initial loss value corresponding to each sample cross cluster.

[0123] In an exemplary embodiment, the above step S320 updates the model parameters of the content recommendation model to be trained according to the initial loss value corresponding to each sample cross cluster, and obtains the updated content recommendation model corresponding to each sample cross cluster, which specifically includes the following contents: determining the initial model parameter update gradient corresponding to each sample cross cluster according to the initial loss value corresponding to each sample cross cluster; obtaining the updated model parameters corresponding to each sample cross cluster according to the current model parameters of the content recommendation model to be trained, the norm of the current model parameters and the initial model parameter update gradient corresponding to each sample cross cluster; updating the current model parameters of the content recommendation model to be trained according to the updated model parameters corresponding to each sample cross cluster, and obtaining the updated content recommendation model corresponding to each sample cross cluster.

[0124] Among them, the initial model parameter update gradient corresponding to each sample cross cluster is obtained by derivatizing the initial loss value corresponding to each sample cross cluster.

[0125] The norm of the current model parameter may refer to the 2-norm of the current model parameter, which is used to reflect the numerical value of the current model parameter. It should be noted that the current model parameter may be considered as a vector.

[0126] Among them, each sample cross cluster corresponds to a set of updated model parameters, which are used to indicate that the initial loss value corresponding to each sample cross cluster reaches the worst specific model parameters, such as making the BPR loss corresponding to each sample cross cluster reach the worst specific model parameters within a certain range of the current model parameters. In addition, by updating the current model parameters of the content recommendation model to be trained to the updated model parameters corresponding to each sample cross cluster, the updated content recommendation model corresponding to each sample cross cluster can be obtained.

[0127] Exemplarily, the server performs a derivative processing on the initial loss value corresponding to each sample cross cluster to obtain the initial model parameter update gradient corresponding to each sample cross cluster; then, based on the current model parameters of the content recommendation model to be trained, the norm of the current model parameters and the initial model parameter update gradient corresponding to each sample cross cluster, the specific model parameters that make the BPR loss corresponding to each sample cross cluster reach the worst within a certain range of the current model parameters are calculated, and used as the updated model parameters corresponding to each sample cross cluster; finally, the current model parameters of the content recommendation model to be trained are updated to the updated model parameters corresponding to each sample cross cluster, and the updated content recommendation model corresponding to each sample cross cluster can be obtained.

[0128] For example, let the cross-bilateral group consisting of the p-th user group and the q-th content group be g p,q , then the BPR loss of the cross-bilateral group is (ie, the initial loss value) is:

[0129]

[0130] in, represents the training data set, including users, content, and user interactions (such as clicks) on the content; θ represents the current model parameters of the content recommendation model to be trained.

[0131] in, represents the cross-bilateral group g p,q The corresponding set of triples; represents the pth user group, represents the qth content group, represents all the content contained in all content groups; u represents the user, i and j both represent the content; y u,i Indicates whether user u interacts with content i, y u,i =1 means user u has interacted with content i; y u,j Indicates whether user u and content j have interaction, yu,j=0 means user u has no interaction with content j. express The size of the set, i.e. The total number of triplets in the set. BPR(·) is the BPR loss of the triplet.

[0132] Among them, the cross-bilateral group g p,q The corresponding sharpness perception loss is defined as the worst loss within a certain range of the current model parameters (such as θ+∈), and T is introduced θ =|θ| to normalize the scales of different parameters to ensure that the scales of different parameters are close:

[0133]

[0134] in, represents the cross-bilateral group g p,q The corresponding sharpness perception loss; ρ represents a hyperparameter used to control the range of the worst loss search; T θ represents the norm of the current model parameters; ∈ represents the offset applied to the current model parameters. This offset should be controlled within a certain range, and its restriction condition is ∈ should change with the magnitude of the overall parameter. When the overall parameter is large, the range should also be larger; It means that the norm of ∈ cannot exceed a certain range.

[0135] Among them, the cross-bilateral group g p,q The specific parameters of the corresponding worst loss can be approximated using the first-order Taylor expansion:

[0136]

[0137] in, Let the crossed bilateral group g p,q BPR loss The worst specific model parameters are achieved; express The corresponding model parameter update gradient is the initial model parameter update gradient. In this way, the parameters corresponding to the worst loss can be directly obtained through Taylor expansion approximation based on the current model parameters, thereby obtaining the worst loss and optimizing the gradient of the worst loss.

[0138] It should be noted that within a certain range of the current model parameters (such as θ+∈), there will always be a maximum value, which can be approximated using Taylor expansion; that is, the cross-bilateral group g p,q The corresponding sharpness perception loss refers to the worst loss among all model parameters within a certain bounded range of the current parameter model.

[0139] The technical solution provided by the embodiment of the present disclosure first determines the initial model parameter update gradient corresponding to each sample cross cluster according to the initial loss value corresponding to each sample cross cluster, and then obtains the updated model parameters corresponding to each sample cross cluster according to the current model parameters of the content recommendation model to be trained, the norm of the current model parameters and the initial model parameter update gradient corresponding to each sample cross cluster. Finally, the current model parameters of the content recommendation model to be trained are updated according to the updated model parameters corresponding to each sample cross cluster, so as to obtain the updated content recommendation model corresponding to each sample cross cluster. In this way, when updating the model, the current model parameters of the content recommendation model to be trained, the norm of the current model parameters and the initial model parameter update gradient corresponding to each sample cross cluster are comprehensively considered, which is conducive to improving the determination accuracy of the updated model parameters corresponding to each sample cross cluster, thereby improving the updating effect of the content recommendation model.

[0140] In an exemplary embodiment, the above step S230, in which the target loss value corresponding to each sample cross cluster is fused to obtain a first loss value, specifically includes the following contents: determining the current weight corresponding to each sample cross cluster; and according to the current weight corresponding to each sample cross cluster, fusing the target loss value corresponding to each sample cross cluster to obtain a first loss value.

[0141] Among them, the current weight corresponding to each sample cross cluster is used to represent the current weight corresponding to the target loss value corresponding to each sample cross cluster, and is dynamically updated.

[0142] Exemplarily, the server first determines the current weight corresponding to each sample cross cluster, and performs weighted summation processing on the target loss value corresponding to each sample cross cluster according to the current weight corresponding to each sample cross cluster to obtain the first loss value. For example, the target loss values ​​corresponding to the sample cross clusters A1B1, A1B2, A2B1 and A2B2 are L1, L2, L3 and L4 respectively, and the corresponding current weights are W1, W2, W3 and W4 respectively, then the first loss value = L1×W1+L2×W2+L3×W3+L4×W4.

[0143] For example, the server calculates the first loss value through the following formula 5, that is, the proxy losses (i.e., sharpness perception losses) of different cross-bilateral groups are re-weighted through the collaborative loss control algorithm:

[0144]

[0145] in, represents the collaborative loss of sharpness perception, i.e., the first loss value; represents the cross-bilateral group g p,q The corresponding sharpness perception loss; wp,q represents the cross-bilateral group g p,q The corresponding current weight.

[0146] The technical solution provided by the embodiment of the present disclosure first determines the current weight corresponding to each sample cross cluster, and then fuses the target loss value corresponding to each sample cross cluster according to the current weight corresponding to each sample cross cluster to obtain a first loss value; in this way, when calculating the first loss value, the current weight corresponding to each sample cross cluster and the target loss value corresponding to each sample cross cluster are comprehensively considered, which can not only improve the calculation accuracy of the first loss value, but also enable the content recommendation model trained based on the first loss value to have similar discrimination capabilities for positive and negative examples of different cross clusters.

[0147] In an exemplary embodiment, the current weight corresponding to each sample cross cluster is determined, specifically including the following contents: according to the target loss value corresponding to each sample cross cluster, the target model parameter update gradient corresponding to each sample cross cluster is determined; according to the target loss value and the target model parameter update gradient corresponding to each sample cross cluster, the cumulative contribution information corresponding to each sample cross cluster is determined; the cumulative contribution information corresponding to each sample cross cluster is used to characterize the fusion processing result of the difference value between the updated target loss value and the target loss value corresponding to each sample cross cluster when the target model parameter update gradient of each sample cross cluster is used to update the content recommendation model to be trained; according to the cumulative contribution information corresponding to each sample cross cluster, the historical weight corresponding to each sample cross cluster is updated to obtain the current weight corresponding to each sample cross cluster.

[0148] Among them, the target model parameter update gradient refers to the sharpness perception gradient; the target model parameter update gradient corresponding to each sample cross cluster refers to the model parameter update gradient of the target loss value corresponding to each sample cross cluster, which can be obtained by differentiating the target loss value corresponding to each sample cross cluster, specifically refers to the sharpness perception gradient of the sharpness perception loss corresponding to each sample cross cluster.

[0149] Among them, the cumulative contribution information corresponding to each sample cross cluster is used to indicate how much the optimization of the sample cross cluster can help the current disadvantaged sample cross cluster, that is, how much the target loss value (i.e., sharpness perception loss) corresponding to each sample cross cluster can be reduced. Moreover, the cumulative contribution information corresponding to each sample cross cluster is calculated based on the target loss value corresponding to each sample cross cluster and the target model parameter update gradient.

[0150] Among them, for a certain sample cross cluster, the content recommendation model to be trained is updated according to the target model parameter update gradient (i.e., sharpness perception gradient) of the sample cross cluster to obtain the updated content recommendation model, and the updated content recommendation model is used to execute steps S220 to 230 again, that is, the sample data group of each sample cross cluster is input into the updated content recommendation model to obtain the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster, and the updated target loss value corresponding to each sample cross cluster is determined according to the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster, and compared with the target loss value (i.e., the original target loss value) corresponding to each sample cross cluster to obtain the difference between the updated target loss value and the target loss value corresponding to each sample cross cluster, and the difference between the updated target loss value and the target loss value corresponding to each sample cross cluster is fused to obtain the fusion result as the cumulative contribution information corresponding to the sample cross cluster.

[0151] The historical weight corresponding to each sample cross cluster refers to the updated weight corresponding to each sample cross cluster during the last model training. For example, during the first model training, the historical weight corresponding to each sample cross cluster is the initialization weight. It is necessary to update each sample cross cluster according to its cumulative contribution information to obtain the updated weight corresponding to each sample cross cluster. During the second model training, the historical weight corresponding to each sample cross cluster refers to the updated weight corresponding to each sample cross cluster during the first model training, and so on.

[0152] The current weight corresponding to each sample cross cluster refers to the updated weight corresponding to each sample cross cluster during the current model training, which is obtained by updating the historical weight corresponding to each sample cross cluster according to the cumulative contribution information corresponding to each sample cross cluster.

[0153] Exemplarily, the server performs derivative processing on the target loss value corresponding to each sample cross cluster to obtain the target model parameter update gradient corresponding to each sample cross cluster; the target loss value and the target model parameter update gradient corresponding to each sample cross cluster are input into the cumulative contribution information calculation model, and the cumulative contribution information corresponding to each sample cross cluster is calculated by the cumulative contribution information calculation model; for example, the server updates the content recommendation model to be trained according to the target model parameter update gradient (i.e., sharpness perception gradient) of each sample cross cluster to obtain the updated content recommendation model corresponding to each sample cross cluster, and uses the updated content recommendation model corresponding to each sample cross cluster to execute steps S220 to 230 again, that is, for each sample cross cluster, each sample cross cluster is respectively updated. The sample data group of the cross cluster is input into the updated content recommendation model corresponding to the sample cross cluster, and the interaction index prediction score corresponding to the updated sample data group of each sample cross cluster is obtained. According to the interaction index prediction score corresponding to the sample data group of each sample cross cluster, the updated target loss value corresponding to each sample cross cluster is determined, and compared with the target loss value corresponding to each sample cross cluster (i.e., the original target loss value), and the difference between the updated target loss value and the target loss value corresponding to each sample cross cluster is obtained. The difference between the updated target loss value and the target loss value corresponding to each sample cross cluster is fused to obtain the fusion result as the cumulative contribution information corresponding to the sample cross cluster, thereby obtaining the cumulative contribution information corresponding to each sample cross cluster. Finally, the server updates the historical weight corresponding to each sample cross cluster according to the cumulative contribution information corresponding to each sample cross cluster, and obtains the updated weight corresponding to each sample cross cluster during this model training as the current weight corresponding to each sample cross cluster.

[0154] The technical solution provided by the embodiment of the present disclosure first determines the target model parameter update gradient corresponding to each sample cross cluster according to the target loss value corresponding to each sample cross cluster, then determines the cumulative contribution information corresponding to each sample cross cluster according to the target loss value and the target model parameter update gradient corresponding to each sample cross cluster, and finally updates the historical weight corresponding to each sample cross cluster according to the cumulative contribution information corresponding to each sample cross cluster to obtain the current weight corresponding to each sample cross cluster; in this way, when determining the current weight corresponding to each sample cross cluster, the current weight corresponding to each sample cross cluster is updated according to the cumulative contribution information corresponding to each sample cross cluster, which is conducive to ensuring the timeliness of the weight corresponding to each sample cross cluster, thereby improving the accuracy of determining the weight corresponding to each sample cross cluster.

[0155] In an exemplary embodiment, the cumulative contribution information corresponding to each sample cross cluster is determined according to the target loss value and the target model parameter update gradient corresponding to each sample cross cluster, which specifically includes the following contents: according to the target model parameter update gradient corresponding to each sample cross cluster, the contribution information of each sample cross cluster to each sample cross cluster is determined; the contribution information of each sample cross cluster to each sample cross cluster is used to characterize the difference between the updated target loss value and the target loss value corresponding to each sample cross cluster when the target model parameter update gradient of each sample cross cluster is used to update the content recommendation model to be trained; according to the target loss value corresponding to each sample cross cluster, the weight of the contribution information of each sample cross cluster to each sample cross cluster is determined; according to the weight of the contribution information of each sample cross cluster to each sample cross cluster, the contribution information of each sample cross cluster to each sample cross cluster is fused to obtain the cumulative contribution information corresponding to each sample cross cluster.

[0156] The contribution information of each sample cross cluster to other sample cross clusters is used to characterize the extent to which the target loss value (i.e., sharpness perception loss) of other sample cross clusters can be reduced when optimizing the sample cross cluster. Moreover, the contribution information of each sample cross cluster to another sample cross cluster is calculated based on the target model parameter update gradient corresponding to the sample cross cluster and the other sample cross cluster.

[0157] Among them, for a certain sample cross cluster, the content recommendation model to be trained is updated according to the target model parameter update gradient (i.e., sharpness perception gradient) of the sample cross cluster to obtain the updated content recommendation model, and the updated content recommendation model is used to execute steps S220 to 230 again, that is, the sample data group of another sample cross cluster is input into the updated content recommendation model to obtain the updated interaction index prediction score corresponding to the sample data group of another sample cross cluster, and the updated target loss value corresponding to the other sample cross cluster is determined according to the updated interaction index prediction score corresponding to the sample data group of another sample cross cluster, and compared with the target loss value corresponding to the other sample cross cluster (i.e., the original target loss value) to obtain the difference between the updated target loss value and the target loss value corresponding to the other sample cross cluster as the contribution information of the sample cross cluster to the other sample cross cluster.

[0158] Among them, the weight of the contribution information of each sample cross cluster to each sample cross cluster is used to characterize the importance of the contribution information of each sample cross cluster to each sample cross cluster, which is specifically calculated through the target loss value corresponding to each sample cross cluster.

[0159] Among them, the cumulative contribution information corresponding to each sample cross cluster is obtained by fusing the contribution information of each sample cross cluster to each sample cross cluster according to the weight of the contribution information of each sample cross cluster to each sample cross cluster, such as weighted summation. Exemplarily, the server can calculate the contribution information of each sample cross cluster to another sample cross cluster according to the target model parameter update gradient corresponding to each sample cross cluster and another sample cross cluster, thereby obtaining the contribution information of each sample cross cluster to each sample cross cluster. For example, the server updates the content recommendation model to be trained according to the target model parameter update gradient (i.e., sharpness perception gradient) of each sample cross cluster, obtains the updated content recommendation model corresponding to each sample cross cluster, and uses the updated content recommendation model corresponding to each sample cross cluster to perform steps S220 to 230 again, that is, for a certain sample cross cluster (such as A1B1), the server inputs the sample data group of another sample cross cluster (such as A1B2) into the updated content recommendation model corresponding to the sample cross cluster (such as A1B1), and obtains the updated content recommendation model corresponding to the sample data group of another sample cross cluster (such as A1B2). According to the interaction index prediction score after the update processing corresponding to the sample data group of another sample cross cluster (such as A1B2), the updated target loss value corresponding to another sample cross cluster (such as A1B2) is determined, and compared with the target loss value (i.e. the original target loss value) corresponding to another sample cross cluster (such as A1B2), and the difference between the updated target loss value and the target loss value corresponding to another sample cross cluster (such as A1B2) is obtained as the contribution information of the sample cross cluster (such as A1B1) to another sample cross cluster (such as A1B2), thereby obtaining the contribution information of each sample cross cluster to each sample cross cluster. Then, the server calculates the weight of the contribution information of each sample cross cluster to each sample cross cluster according to the target loss value corresponding to each sample cross cluster; finally, according to the weight of the contribution information of each sample cross cluster to each sample cross cluster, the contribution information of each sample cross cluster to each sample cross cluster is weighted and summed to obtain the cumulative contribution information corresponding to each sample cross cluster. For example, the contribution information of sample cross cluster A1B1 to sample cross clusters A1B1, A1B2, A2B1 and A2B2 are G1, G2, G3 and G4 respectively, and the weights corresponding to each contribution information are W11, W12, W13 and W14 respectively, then the cumulative contribution information of sample cross cluster A1B1 to sample cross clusters A1B1, A1B2, A2B1 and A2B2 = G1×W11+G2×W12+G3×W13+G4×W14.

[0160] For example, the server can calculate the cross-bilateral group g through the following formula 6: p,q The corresponding weight w p,q ; First, the server estimates the contribution of each cross-bilateral group to the other cross-bilateral group:

[0161]

[0162] Among them, C(g p,q → a,b ) represents the cross-bilateral population g p,q For another cross-bilateral group g a,b The contribution information of the optimized cross-bilateral group g p,q When the cross-bilateral population g is reduced, a,b The sharpness perception loss is to use the cross-bilateral group g p,q The loss of sharpness perception The sharpness perception gradient Update the model parameters of the content recommendation model, and use the updated content recommendation model to calculate the cross-bilateral group g a,b The new sharpness-aware loss is combined with the cross-bilateral group g a,b The original perceptual loss of sharpness is compared.

[0163] Among them, α represents a hyperparameter, which is used to indicate the step size of gradient descent; Indicates the loss of sharpness perception The sharpness perception gradient, i.e., the target model parameter update gradient; represents the recalculated cross-bilateral population g a,b loss of perceived sharpness; Represents the sharpness perception gradient The transpose of , which is the horizontal quantity; Indicates the loss of sharpness perception The sharpness-perceived gradient of , which is a column vector.

[0164] Next, the server calculates a cross-bilateral group g according to the following formula 7: p,q The cumulative contribution information to fairness can be intuitively understood as the extent to which the optimization of the cross-bilateral group can help the current disadvantaged group.

[0165]

[0166] Among them, C(g p,q ) represents the cross-bilateral group g p,q The cumulative contribution information, β a,b represents the cross-bilateral group g p,q For another cross-bilateral group ga,b Contribution information C(g p,q → a,b ) weight; γ represents a hyperparameter.

[0167] Finally, based on the cumulative contribution information, the server can optimize the cross-bilateral population g after each batch of gradient descent. p,q The weight w p,q Update, that is, the server uses exp(η·C(g p,q )) for the crossed bilateral group g p,q The last weight w p,q (i.e. historical weight) is updated to obtain w p,q ·exp(η·C(g p,q )), and w p,q ·exp(η·C(g p,q )) is normalized to obtain the cross-bilateral population g p,q The weight of this time is w p,q (i.e. current weight):

[0168]

[0169] Where η represents a hyperparameter; C(g a,b ) represents the cross-bilateral group g a,b Cumulative contribution information of w a,b represents the cross-bilateral group g a,b The previous weight (i.e. historical weight).

[0170] The technical solution provided by the embodiment of the present disclosure first updates the gradient of the target model parameters corresponding to each sample cross cluster to determine the contribution information of each sample cross cluster to each sample cross cluster, then determines the weight of the contribution information of each sample cross cluster to each sample cross cluster according to the target loss value corresponding to each sample cross cluster, and finally, based on the weight of the contribution information of each sample cross cluster to each sample cross cluster, fuses the contribution information of each sample cross cluster to each sample cross cluster to obtain the cumulative contribution information corresponding to each sample cross cluster; in this way, when calculating the cumulative contribution information corresponding to each sample cross cluster, the contribution information of each sample cross cluster to each sample cross cluster and the weight of the contribution information of each sample cross cluster to each sample cross cluster are comprehensively considered, which is conducive to improving the calculation accuracy of the cumulative contribution information corresponding to each sample cross cluster, and further improves the determination accuracy of the current weight obtained based on the cumulative contribution information.

[0171] In an exemplary embodiment, the above-mentioned step S240 determines the second loss value according to the first interaction indicator prediction score corresponding to the first sample data of each sample cross-cluster, and specifically includes the following contents: determining the average interaction indicator prediction score of each sample cross-cluster according to the first interaction indicator prediction score corresponding to the first sample data of each sample cross-cluster; determining the second loss value according to the average interaction indicator prediction score of each sample cross-cluster.

[0172] Among them, the average interaction index prediction score of each sample cross-cluster refers to the average value of the first interaction index prediction scores corresponding to the first sample data (i.e., positive examples) of each sample cross-cluster; for example, the first interaction index prediction scores corresponding to the first sample data of a sample cross-cluster are F1, F2, and F3 respectively, then the average interaction index prediction score of the sample cross-cluster is (F1+F2+F3) / 3.

[0173] The second loss value is calculated based on the average interaction index prediction score of each sample cross cluster.

[0174] Exemplarily, the server averages the first interaction indicator prediction scores corresponding to the first sample data of each sample cross-cluster to obtain the average interaction indicator prediction score of each sample cross-cluster; and calculates the second loss value based on the average interaction indicator prediction score of each sample cross-cluster and the sub-group alignment loss function.

[0175] For example, the server uses subgroup alignment loss and prediction score normalization to align the prediction scores of positive interaction indicators of different cross-bilateral groups to prevent positive examples of a certain type of cross-bilateral group from being systematically discriminated against. First, a subgroup alignment loss is introduced to explicitly enforce the consistency of the prediction scores:

[0176]

[0177] in, Represents the subgroup alignment loss, that is, the second loss value; represents the cross-bilateral group g p,j The average interaction metric prediction score of represents the cross-bilateral group g p,k The average interaction index prediction score of ; P represents the number of user groups, and Q represents the number of content groups. It should be noted that here we only need to align different cross-bilateral groups under the same user group, because the interaction index prediction scores between different user groups will not be compared with each other.

[0178] In addition, the server can also use interaction indicator prediction score normalization to constrain the range of interaction indicator prediction scores and implicitly enhance the consistency of interaction indicator prediction scores. Specifically, during the model training phase, when the content recommendation model outputs the corresponding interaction indicator prediction score for user-content interaction, the traditional click prediction score can be normalized. Normalization is as follows:

[0179]

[0180] in, It represents the interaction index prediction score of the positive example of user-content interaction, that is, the first interaction index prediction score; u represents the account feature vector, v represents the content feature vector, both of which can be obtained through different recommendation representation models (such as MF, LightGCN, etc.); τ represents the hyperparameter.

[0181] The technical solution provided by the embodiment of the present disclosure first determines the average interaction index prediction score of each sample cross-cluster based on the first interaction index prediction score corresponding to the first sample data of each sample cross-cluster, and then determines the second loss value based on the average interaction index prediction score of each sample cross-cluster; in this way, when calculating the second loss value, the average interaction index prediction score of the first interaction index prediction score of each sample cross-cluster is comprehensively considered, which can not only improve the calculation accuracy of the second loss value, but also enable the content recommendation model trained based on the second loss value to align the interaction index prediction scores of positive examples of different cross-clusters, so as to prevent the positive example of a certain cross-cluster from being systematically ranked behind the positive example of another cross-cluster, which is conducive to exploring the potential interests of cross-clusters, thereby making the output results of the trained content recommendation model more accurate.

[0182] Figure 4 is a flow chart of a content recommendation method according to an exemplary embodiment. Figure 4 As shown, the content recommendation method is used in a server and specifically includes the following steps:

[0183] In step S410, the content to be recommended corresponding to the account to be recommended is obtained.

[0184] In step S420, the account features of the account to be recommended and the content features of the content to be recommended are input into the trained content recommendation model for prediction processing to obtain the interaction indicator prediction score of the content to be recommended; the trained content recommendation model is trained by the content recommendation model training method.

[0185] In step S430, target content is screened out from the content to be recommended according to the interaction indicator prediction score, and the target content is recommended to the account to be recommended.

[0186] The content to be recommended corresponding to the account to be recommended may refer to content on the Internet or candidate content, such as pictures, videos, articles, live broadcasts, etc.

[0187] Among them, the interaction index prediction score of the content to be recommended refers to the interaction index prediction score of the recommended account towards the recommended content, which is used to indicate the prediction score of the interactive operation of the recommended account towards the recommended content, such as the click prediction score, the like prediction score, etc.

[0188] The target content refers to content with a high predicted interaction index score, or content with a predicted interaction index score greater than a preset score.

[0189] Exemplarily, the server obtains the content to be recommended corresponding to the account to be recommended from the network, and then performs feature extraction processing on the account information of the account to be recommended and the content information of the content to be recommended to obtain the account features of the account to be recommended and the content features of the content to be recommended; then the account features of the account to be recommended and the content features of the content to be recommended are input into the trained content recommendation model, and the account features of the account to be recommended and the content features of the content to be recommended are subjected to a series of processing (including normalization processing) through the trained content recommendation model to obtain the interaction index prediction score of the content to be recommended; finally, the content to be recommended is sorted in descending order of the interaction index prediction score to obtain the sorted content to be recommended; the top N content (N is a positive integer greater than or equal to 2) is selected from the sorted content to be recommended as the target content, and the target content is recommended to the account to be recommended.

[0190] In the above content recommendation method, the content to be recommended corresponding to the account to be recommended is first obtained, and then the interaction index prediction score of the content to be recommended is output through the trained content recommendation model; finally, the target content is screened out from the content to be recommended according to the interaction index prediction score, and the target content is recommended to the account to be recommended. In this way, when performing content recommendation, the content to be recommended corresponding to the account to be recommended is first obtained, and then the target content is screened out from the content to be recommended according to the interaction index prediction score output by the content recommendation model, and recommended to the account to be recommended, which is conducive to improving the accuracy of content recommendation. Moreover, the content recommendation model is trained based on the first loss value and the second loss value, and it can have similar ability to distinguish positive and negative examples of different cross clusters, and can align the interaction index prediction scores of positive examples of different cross clusters, so as to prevent the positive example of a cross cluster from being systematically ranked behind the positive example of another cross cluster, which is conducive to exploring the potential interests of the cross cluster, thereby making the output result of the trained content recommendation model more accurate, and further improving the accuracy of content recommendation.

[0191] Figure 5is a flowchart of another content recommendation method according to an exemplary embodiment. Figure 5 As shown, the content recommendation method is used in a server and specifically includes the following steps:

[0192] In step S501, sample cross-clustering and sample data groups of the sample cross-clustering are obtained.

[0193] Among them, each sample cross-cluster includes a sample account cluster and a sample content cluster; each sample data group includes first sample data and second sample data; the first sample data includes a sample account in the sample account cluster in the sample cross-cluster to which it belongs, and a sample content with which the sample account has interacted in the sample content cluster in the sample cross-cluster to which it belongs; the second sample data includes the sample account and a sample content with which the sample account has not interacted in each sample cross-cluster.

[0194] In step S502, the sample data group of each sample cross-cluster is input into the content recommendation model to be trained to obtain the interaction indicator prediction score corresponding to the sample data group of each sample cross-cluster; the interaction indicator prediction score includes the first interaction indicator prediction score corresponding to the first sample data, and the second interaction indicator prediction score corresponding to the second sample data.

[0195] In step S503, the loss value corresponding to the sample data group of each sample cross cluster is determined according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster; the loss values ​​corresponding to the sample data group of each sample cross cluster are fused respectively to obtain the initial loss value corresponding to each sample cross cluster.

[0196] In step S504, the initial model parameter update gradient corresponding to each sample cross cluster is determined according to the initial loss value corresponding to each sample cross cluster; the updated model parameters corresponding to each sample cross cluster are obtained according to the current model parameters of the content recommendation model to be trained, the norm of the current model parameters and the initial model parameter update gradient corresponding to each sample cross cluster.

[0197] In step S505, the current model parameters of the content recommendation model to be trained are updated according to the updated model parameters corresponding to each sample cross cluster, so as to obtain an updated content recommendation model corresponding to each sample cross cluster.

[0198] In step S506, the updated content recommendation model corresponding to each sample cross cluster is used to output the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster; the updated loss value corresponding to the sample data group of each sample cross cluster is determined according to the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster, and the updated loss value corresponding to the sample data group of each sample cross cluster is fused to obtain the target loss value corresponding to each sample cross cluster.

[0199] In step S507, the target model parameter update gradient corresponding to each sample cross cluster is determined according to the target loss value corresponding to each sample cross cluster; and the contribution information of each sample cross cluster to each sample cross cluster is determined according to the target model parameter update gradient corresponding to each sample cross cluster.

[0200] Among them, the contribution information of each sample cross-cluster to each sample cross-cluster is used to characterize the difference between the updated target loss value and the target loss value corresponding to each sample cross-cluster when the target model parameter update gradient of each sample cross-cluster is used to update the content recommendation model to be trained.

[0201] In step S508, the weight of the contribution information of each sample cross cluster to each sample cross cluster is determined according to the target loss value corresponding to each sample cross cluster.

[0202] In step S509 , according to the weight of the contribution information of each sample cross cluster to each sample cross cluster, the contribution information of each sample cross cluster to each sample cross cluster is fused to obtain the cumulative contribution information corresponding to each sample cross cluster.

[0203] In step S510, the historical weight corresponding to each sample cross cluster is updated according to the cumulative contribution information corresponding to each sample cross cluster, so as to obtain the current weight corresponding to each sample cross cluster.

[0204] In step S511, according to the current weight corresponding to each sample cross cluster, the target loss value corresponding to each sample cross cluster is fused to obtain a first loss value.

[0205] In step S512, the average interaction index prediction score of each sample cross cluster is determined according to the first interaction index prediction score corresponding to the first sample data of each sample cross cluster; and the second loss value is determined according to the average interaction index prediction score of each sample cross cluster.

[0206] In step S513, the content recommendation model to be trained is iteratively trained according to the first loss value and the second loss value to obtain a trained content recommendation model.

[0207] In step S514, the content to be recommended corresponding to the account to be recommended is obtained; and the interaction index prediction score of the content to be recommended is output through the trained content recommendation model.

[0208] In step S515, target content is screened out from the content to be recommended according to the interaction indicator prediction score, and the target content is recommended to the account to be recommended.

[0209] In the above content recommendation method, in the process of training the content recommendation model, modeling is performed for sample cross-clusters, and the first loss value constructed based on the interaction index prediction score corresponding to the sample data group of each sample cross-cluster is comprehensively considered, so that the trained content recommendation model can have similar discrimination ability for positive and negative examples of different cross-clusters; at the same time, the second loss value constructed based on the first interaction index prediction score corresponding to the first sample data of each sample cross-cluster is considered, so that the trained content recommendation model can align the interaction index prediction scores of positive examples of different cross-clusters to prevent the positive examples of a certain cross-cluster from being systematically ranked behind the positive examples of another cross-cluster, which is conducive to exploring the potential interests of cross-clusters, thereby making the output results of the trained content recommendation model more accurate, thereby improving the accuracy of content recommendation.

[0210] In order to more clearly illustrate the content recommendation model training method provided by the embodiment of the present disclosure, the content recommendation model training method is specifically described below with a specific embodiment. Figure 6 As shown, the present disclosure also provides a training method for cross-bilateral fair recommendation. Specifically, the present disclosure hopes that the recommendation algorithm can fairly model the potential interests of different cross-bilateral groups. To this end, the present disclosure designs a collaborative loss control optimization method that takes sharpness into consideration to control the positive and negative example discrimination capabilities of different cross-bilateral groups, and uses explicit alignment loss and implicit score normalization to align the positive example prediction scores of different groups, that is, the positive example prediction scores of different groups are aligned through sub-group alignment loss, thereby preventing the positive examples of a certain group from being systematically ranked behind the positive examples of another group, thereby effectively alleviating the problem of cross-bilateral unfairness without significantly losing recommendation accuracy.

[0211] The present disclosure hopes that the recommendation algorithm can fairly model the potential interests of different cross-bilateral groups, that is, for potential interests belonging to different groups (positive examples of user-content interaction), they can all be ranked at the top in the recommendation list. Specifically, cross-bilateral fairness is divided into two tasks: controlling the ability to distinguish between positive and negative examples of different cross-groups and aligning the predicted ranking scores of different cross-groups. The former is achieved through a sharpness-aware collaborative loss control algorithm, and the latter is achieved through explicit alignment loss and implicit prediction score normalization, that is, the present disclosure is designed based on cross-bilateral groups.

[0212] refer to Figure 6 , each row represents a user, and each column corresponds to a content. Therefore, each icon in the content box to be recommended is a (u, i) pair, and there will be an interaction prediction score, such as d1-d6. Users may interact with it, that is, they are interested in it, and if they are not interested, it is a cross; in this legend, there are two groups for both users and content, so there are four cross-bilateral groups, namely male users-horror movies, male users-love movies, female users-horror movies, and female users-love movies; it should be noted that the cross is not considered here, because the present disclosure is concerned with the user's interests, that is, positive examples. Take d1 as an example, it means that within this user, the interaction prediction score of the first content minus the interaction prediction score of the second content. For goal 1, the present disclosure hopes that the content recommendation model has the same ability to distinguish positive and negative examples for different cross-bilateral groups, so the difference in the positive and negative example scores of different cross-bilateral groups should be close, that is, d1≈d2≈d4≈d6. The following goal 2 is similar. The present disclosure hopes that the interaction prediction scores of positive examples in different cross-bilateral groups are close, that is, d3≈d5≈0. Goal 1 and Goal 2 here correspond to Step 1 and Step 2 below respectively.

[0213] Step 1: Sharpness-aware collaborative loss control

[0214] The purpose of this step is to hope that the content recommendation model obtained by training can have similar distinguishing ability for positive and negative examples of different cross-bilateral groups. The present disclosure uses a re-weighted method to control the distinguishing ability. To this end, this step first requires an optimizable agent for the ability to distinguish between positive and negative examples in a group. Compared with directly using training loss as a proxy for test ability, sharpness perception takes into account the geometric properties of the loss function and has better training and testing consistency. To this end, the present disclosure proposes a sharpness-aware cross-bilateral group loss, which is specifically explained by taking the loss function BPR, which is often used in Top-K recommendations, as an example. The present disclosure can also use other common recommendation losses, such as cross entropy loss, Hinge loss, etc. Please refer to the above formula 2-formula 8 and related content for details.

[0215] Step 2: Positive score alignment

[0216] The present disclosure uses subgroup alignment loss and prediction score normalization to align the positive example prediction scores of different cross-bilateral groups to prevent a certain type of positive examples from being systematically discriminated against. For details, refer to the above formula 9-formula 10 and related content.

[0217] In addition, the overall training process of this method is: the input includes the training data set P×Q cross-bilateral groups consisting of P user groups and Q content groups, learning rate lr, hyperparameters λ, α, τ, η, γ, ρ, untrained content recommendation model f(θ); output includes the trained content recommendation model f(θ). The specific training process is as follows:

[0218] (1) Initialize the content recommendation model f(θ) and group weights For p=1,…,P and q=1,…,Q.

[0219] (2) Train for a specified epoch according to (3)-(14).

[0220] (3) From the dataset Randomly select a batch of data

[0221] (4) For p=1 to P user groups, perform the following training (5)-(9).

[0222] (5) For q=1 to Q content groups, perform the following training (6)-(8).

[0223] (6) Use the content recommendation model f(θ) to make predictions, use prediction normalization, calculate the recommendation loss (such as BPR loss) for the predicted value, and calculate the cross-bilateral group g p,q Gradient

[0224] The server first uses the content prediction model f(θ) to target the cross-bilateral group g p,q The (u,i) and (u,j) in the triple in , calculate the corresponding interaction prediction score (such as click prediction score), and then perform normalization to obtain the normalized interaction prediction score; calculate the BPR loss BPR(u,i,j) of the triple according to the normalized interaction prediction score; according to the cross bilateral group g p,q The BPR loss BPR(u,i,j) of all triples is used to calculate the BPR loss of the cross-bilateral group. Then according to the BPR loss of the cross-bilateral group Calculate the crossover bilateral population g p,q Gradient

[0225] (7) According to the calculated cross-bilateral population g p,q Gradient Calculate the corresponding sharpness perception loss (It also includes prediction normalization, which will not be repeated below).

[0226] Among them, the server is based on θ, T θ and gradient Calculate the specific parameters of the worst loss Then Substituting into formula 2, we can recalculate and will recalculate the As a loss of sharpness perception Similarly, the server can calculate the cross-bilateral group g a,b The loss of sharpness perception

[0227] (8) According to the sharpness perception loss, calculate the corresponding sharpness perception gradient ( and ).

[0228] Among them, the server perceives loss based on sharpness Calculate the corresponding sharpness perception gradient And according to the loss of sharpness perception Calculate the corresponding sharpness perception gradient

[0229] (9) End the loop of (5).

[0230] (10) End the loop of (4).

[0231] (11) According to the obtained sharpness perception gradient and sharpness perception loss, calculate g p,q cumulative contribution.

[0232] Among them, the server perceives loss based on sharpness and Calculating Beta a,b ; According to the gradient Calculate the corresponding transpose and get And according to α, Calculate the crossover population g p,q For another crossover group g a,b The contribution of C(g p,q → a,b ), and finally according to C(g p,q → a,b ) and β a,b , calculate the crossover bilateral population g p,q Cumulative contribution to fairness C(g p,q ).

[0233] (12) Update the group weights w for all groups based on the cumulative contribution p,q .

[0234] The server uses exp(η·C(g p,q )) For the last w p,q Update and get w p,q ·exp(η·C(g p,q )),to w p,q ·exp(η·(g p,q )) is normalized to obtain w p,q .

[0235] (13) Calculate the total loss based on the group weight Perform a gradient descent update.

[0236] Among them, the server perceives loss based on sharpness and group weight w p,q , calculate the collaborative loss of sharpness perception In addition, for the pj group, the positive example interaction prediction score of the group is calculated and the average is taken to obtain For the pk group, calculate the positive example interaction prediction score of the group and take the average value to get Then and Substitute into formula 9 and calculate the subgroup alignment loss Finally, according to and Calculate the total loss value And according to the total loss value Train the content recommendation model f(θ).

[0237] (14) End the loop (3).

[0238] (15) End the loop (2).

[0239] (16) Output the trained content recommendation model f(θ).

[0240] The above-mentioned cross-bilateral fair recommendation training method proposes to improve cross-bilateral fairness through sharpness-aware collaborative loss control and prediction score alignment, which can significantly alleviate the problem of cross-bilateral unfairness without significantly losing recommendation accuracy. At the same time, this method can be applied to a variety of recommendation representation models and is compatible with fairness reranking methods.

[0241] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0242] It can be understood that the same / similar parts between the various embodiments of the above method in this specification can refer to each other, and each embodiment focuses on the differences from other embodiments. For related points, please refer to the description of other method embodiments.

[0243] Based on the same inventive concept, the embodiment of the present disclosure also provides a content recommendation model training device for implementing the above-mentioned content recommendation model training method.

[0244] Figure 7 is a block diagram of a content recommendation model training device according to an exemplary embodiment. Figure 7 , the device includes a data acquisition unit 710, a score determination unit 720, a first calculation unit 730, a second calculation unit 740 and a model training unit 750.

[0245] The data acquisition unit 710 is configured to execute acquisition of sample cross-cluster and sample data groups of sample cross-cluster; each sample cross-cluster includes a sample account cluster and a sample content cluster; each sample data group includes first sample data and second sample data; the first sample data includes a sample account in the sample account cluster in the sample cross-cluster to which it belongs, and a sample content with which the sample account has interacted in the sample content cluster in the sample cross-cluster to which it belongs; the second sample data includes the sample account and a sample content with which the sample account has not interacted in each sample cross-cluster.

[0246] The score determination unit 720 is configured to input the sample data group of each sample cross-cluster into the content recommendation model to be trained to obtain the interaction indicator prediction score corresponding to the sample data group of each sample cross-cluster; the interaction indicator prediction score includes the first interaction indicator prediction score corresponding to the first sample data, and the second interaction indicator prediction score corresponding to the second sample data.

[0247] The first calculation unit 730 is configured to execute the interaction indicator prediction score corresponding to the sample data group of each sample cross cluster, determine the target loss value corresponding to each sample cross cluster, and fuse the target loss value corresponding to each sample cross cluster to obtain a first loss value.

[0248] The second calculation unit 740 is configured to determine a second loss value according to the first interaction indicator prediction score corresponding to the first sample data of each sample cross cluster.

[0249] The model training unit 750 is configured to perform iterative training on the content recommendation model to be trained according to the first loss value and the second loss value to obtain a trained content recommendation model.

[0250] In an exemplary embodiment, the first computing unit 730 is further configured to execute, according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster, to determine the initial loss value corresponding to each sample cross cluster; according to the initial loss value corresponding to each sample cross cluster, respectively, to update the model parameters of the content recommendation model to be trained, to obtain the updated content recommendation model corresponding to each sample cross cluster; respectively, through the updated content recommendation model corresponding to each sample cross cluster, to output the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster; according to the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster, to determine the updated loss value corresponding to the sample data group of each sample cross cluster, respectively, to fuse the updated loss values ​​corresponding to the sample data group of each sample cross cluster, to obtain the target loss value corresponding to each sample cross cluster.

[0251] In an exemplary embodiment, the first computing unit 730 is further configured to execute the interaction indicator prediction score corresponding to the sample data group of each sample cross cluster, determine the loss value corresponding to the sample data group of each sample cross cluster; and perform fusion processing on the loss values ​​corresponding to the sample data group of each sample cross cluster respectively to obtain the initial loss value corresponding to each sample cross cluster.

[0252] In an exemplary embodiment, the first computing unit 730 is further configured to determine the initial model parameter update gradient corresponding to each sample cross cluster according to the initial loss value corresponding to each sample cross cluster; obtain the updated model parameters corresponding to each sample cross cluster according to the current model parameters of the content recommendation model to be trained, the norm of the current model parameters and the initial model parameter update gradient corresponding to each sample cross cluster; update the current model parameters of the content recommendation model to be trained according to the updated model parameters corresponding to each sample cross cluster, and obtain the updated content recommendation model corresponding to each sample cross cluster.

[0253] In an exemplary embodiment, the first computing unit 730 is further configured to determine the current weight corresponding to each sample cross cluster; and perform fusion processing on the target loss value corresponding to each sample cross cluster according to the current weight corresponding to each sample cross cluster to obtain a first loss value.

[0254] In an exemplary embodiment, the first computing unit 730 is further configured to determine the target model parameter update gradient corresponding to each sample cross cluster according to the target loss value corresponding to each sample cross cluster; determine the cumulative contribution information corresponding to each sample cross cluster according to the target loss value and the target model parameter update gradient corresponding to each sample cross cluster; the cumulative contribution information corresponding to each sample cross cluster is used to characterize the fusion processing result of the difference value between the updated target loss value and the target loss value corresponding to each sample cross cluster when the target model parameter update gradient of each sample cross cluster is used to update the content recommendation model to be trained; according to the cumulative contribution information corresponding to each sample cross cluster, the historical weight corresponding to each sample cross cluster is updated to obtain the current weight corresponding to each sample cross cluster.

[0255] In an exemplary embodiment, the first computing unit 730 is further configured to execute, according to the target model parameter update gradient corresponding to each sample cross cluster, to determine the contribution information of each sample cross cluster to each sample cross cluster; the contribution information of each sample cross cluster to each sample cross cluster is used to characterize the difference between the updated target loss value and the target loss value corresponding to each sample cross cluster when the content recommendation model to be trained is updated using the target model parameter update gradient of each sample cross cluster; according to the target loss value corresponding to each sample cross cluster, the weight of the contribution information of each sample cross cluster to each sample cross cluster is determined; according to the weight of the contribution information of each sample cross cluster to each sample cross cluster, the contribution information of each sample cross cluster to each sample cross cluster is fused to obtain the cumulative contribution information corresponding to each sample cross cluster.

[0256] In an exemplary embodiment, the second computing unit 740 is further configured to determine the average interaction index prediction score of each sample cross cluster based on the first interaction index prediction score corresponding to the first sample data of each sample cross cluster; and determine the second loss value based on the average interaction index prediction score of each sample cross cluster.

[0257] Based on the same inventive concept, an embodiment of the present disclosure further provides a content recommendation device for implementing the above-mentioned content recommendation method.

[0258] Figure 8is a block diagram of a content recommendation device according to an exemplary embodiment. Figure 8 , the device includes a content acquisition unit 810, a score output unit 820 and a content recommendation unit 830.

[0259] The content acquisition unit 810 is configured to acquire the content to be recommended corresponding to the account to be recommended.

[0260] The score output unit 820 is configured to execute prediction processing by inputting the account characteristics of the account to be recommended and the content characteristics of the content to be recommended into the trained content recommendation model to obtain the interaction indicator prediction score of the content to be recommended; the trained content recommendation model is trained by the content recommendation model training method.

[0261] The content recommendation unit 830 is configured to filter out target content from the content to be recommended according to the interaction indicator prediction score, and recommend the target content to the account to be recommended.

[0262] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0263] Each module in the above-mentioned content recommendation model training device or content recommendation device can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0264] Fig. 9 1 is a block diagram of an electronic device 900 for implementing a content recommendation model training method or a content recommendation method according to an exemplary embodiment. For example, the electronic device 900 may be a server. Fig. 9 , the electronic device 900 includes a processing component 920, which further includes one or more processors, and a memory resource represented by a memory 922 for storing instructions executable by the processing component 920, such as an application. The application stored in the memory 922 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 920 is configured to execute the instructions to perform the above method.

[0265] The electronic device 900 may further include a power supply component 924 configured to perform power management of the electronic device 900, a wired or wireless network interface 926 configured to connect the electronic device 900 to a network, and an input / output (I / O) interface 928. The electronic device 900 may operate based on an operating system stored in the memory 922, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, or the like.

[0266] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory 922 including instructions, and the above instructions can be executed by a processor of the electronic device 900 to perform the above method. The storage medium can be a computer-readable storage medium, for example, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0267] In an exemplary embodiment, a computer program product is further provided. The computer program product includes instructions, and the instructions can be executed by a processor of the electronic device 900 to complete the above method.

[0268] It should be noted that the above-mentioned devices, electronic devices, computer-readable storage media, computer program products, etc. may also include other implementation methods according to the description of the method embodiments. The specific implementation methods can refer to the description of the relevant method embodiments, which will not be described one by one here.

[0269] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.

[0270] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A content recommendation model training method, characterized in that: include: Obtain sample cross clusters and sample data groups of the sample cross clusters; each sample cross cluster includes a sample account cluster and a sample content cluster; each sample data group includes first sample data and second sample data; the first sample data includes a sample account in the sample account cluster in the sample cross cluster to which it belongs, and a sample content with which the sample account has interacted in the sample content cluster in the sample cross cluster to which it belongs; the second sample data includes the sample account and a sample content with which the sample account has not interacted in each of the sample cross clusters; Inputting the sample data group of each sample cross cluster into the content recommendation model to be trained, and obtaining the interaction index prediction score corresponding to the sample data group of each sample cross cluster; the interaction index prediction score includes the first interaction index prediction score corresponding to the first sample data, and the second interaction index prediction score corresponding to the second sample data; Determine the target loss value corresponding to each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster, and perform fusion processing on the target loss value corresponding to each sample cross cluster to obtain a first loss value; Determining a second loss value according to a first interaction indicator prediction score corresponding to the first sample data of each sample cross cluster; The content recommendation model to be trained is iteratively trained according to the first loss value and the second loss value to obtain a trained content recommendation model.

2. The method according to claim 1, characterized in that The step of determining the target loss value corresponding to each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster comprises: Determine the initial loss value corresponding to each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster; According to the initial loss value corresponding to each sample cross cluster, the model parameters of the content recommendation model to be trained are updated to obtain an updated content recommendation model corresponding to each sample cross cluster; Outputting the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster through the updated content recommendation model corresponding to each sample cross cluster; According to the updated interaction index prediction score corresponding to the sample data group of each sample cross cluster, the updated loss value corresponding to the sample data group of each sample cross cluster is determined, and the updated loss values ​​corresponding to the sample data group of each sample cross cluster are fused respectively to obtain the target loss value corresponding to each sample cross cluster.

3. The method according to claim 2, characterized in that The determining, according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster, the initial loss value corresponding to each sample cross cluster comprises: Determine the loss value corresponding to the sample data group of each sample cross cluster according to the interaction index prediction score corresponding to the sample data group of each sample cross cluster; The loss values ​​corresponding to the sample data groups of each sample cross cluster are fused respectively to obtain the initial loss value corresponding to each sample cross cluster.

4. The method according to claim 2, characterized in that: The updating of the model parameters of the content recommendation model to be trained according to the initial loss value corresponding to each sample cross cluster to obtain the updated content recommendation model corresponding to each sample cross cluster includes: Determine the initial model parameter update gradient corresponding to each sample cross cluster according to the initial loss value corresponding to each sample cross cluster; According to the current model parameters of the content recommendation model to be trained, the norm of the current model parameters and the initial model parameters corresponding to each sample cross cluster, the gradient is updated to obtain the updated model parameters corresponding to each sample cross cluster; According to the updated model parameters corresponding to each sample cross cluster, the current model parameters of the content recommendation model to be trained are updated to obtain the updated content recommendation model corresponding to each sample cross cluster.

5. The method according to claim 1, characterized in that The fusing the target loss values ​​corresponding to each sample cross cluster to obtain a first loss value includes: Determine the current weight corresponding to each sample cross cluster; According to the current weight corresponding to each sample cross cluster, the target loss value corresponding to each sample cross cluster is fused to obtain a first loss value.

6. The method according to claim 5, characterized in that The determining of the current weight corresponding to each sample cross cluster includes: Determine, according to the target loss value corresponding to each sample cross cluster, the target model parameter update gradient corresponding to each sample cross cluster; Determine the cumulative contribution information corresponding to each sample cross cluster according to the target loss value and the target model parameter update gradient corresponding to each sample cross cluster; the cumulative contribution information corresponding to each sample cross cluster is used to characterize the fusion processing result of the difference value between the updated target loss value corresponding to each sample cross cluster and the target loss value when the target model parameter update gradient of each sample cross cluster is used to update the content recommendation model to be trained; According to the accumulated contribution information corresponding to each sample cross cluster, the historical weight corresponding to each sample cross cluster is updated to obtain the current weight corresponding to each sample cross cluster.

7. The method according to claim 6, characterized in that The step of updating the target loss value and the target model parameter gradient corresponding to each sample cross cluster to determine the cumulative contribution information corresponding to each sample cross cluster includes: Determine the contribution information of each sample cross cluster to each of the sample cross clusters according to the target model parameter update gradient corresponding to each sample cross cluster; the contribution information of each sample cross cluster to each of the sample cross clusters is used to characterize the difference between the updated target loss value and the target loss value corresponding to each of the sample cross clusters when the content recommendation model to be trained is updated using the target model parameter update gradient of each sample cross cluster; Determining, according to the target loss value corresponding to each sample cross cluster, the weight of the contribution information of each sample cross cluster to each of the sample cross clusters; According to the weight of the contribution information of each sample cross cluster to each sample cross cluster, the contribution information of each sample cross cluster to each sample cross cluster is fused to obtain the cumulative contribution information corresponding to each sample cross cluster.

8. The method according to any one of claims 1 to 7, characterized in that: The determining the second loss value according to the first interaction indicator prediction score corresponding to the first sample data of each sample cross cluster includes: Determining an average interaction index prediction score of each sample cross cluster according to the first interaction index prediction score corresponding to the first sample data of each sample cross cluster; The second loss value is determined according to the average interaction index prediction score of each sample cross cluster.

9. A content recommendation method, characterized in that: include: Obtain the recommended content corresponding to the account to be recommended; Inputting the account features of the account to be recommended and the content features of the content to be recommended into a trained content recommendation model for prediction processing to obtain a predicted score of the interaction index of the content to be recommended; the trained content recommendation model is obtained by training the content recommendation model training method according to any one of claims 1 to 8; Filter out target content from the content to be recommended according to the interaction indicator prediction score, and recommend the target content to the account to be recommended.

10. A content recommendation model training device, characterized in that: include: A data acquisition unit is configured to acquire sample cross clusters and sample data groups of the sample cross clusters; each sample cross cluster includes a sample account cluster and a sample content cluster; each sample data group includes first sample data and second sample data; the first sample data includes a sample account in the sample account cluster in the sample cross cluster to which it belongs, and a sample content with which the sample account has interacted in the sample content cluster in the sample cross cluster to which it belongs; the second sample data includes the sample account and a sample content with which the sample account has not interacted in each of the sample cross clusters; A score determination unit is configured to input the sample data group of each sample cross cluster into the content recommendation model to be trained, and obtain the interaction index prediction score corresponding to the sample data group of each sample cross cluster; the interaction index prediction score includes a first interaction index prediction score corresponding to the first sample data, and a second interaction index prediction score corresponding to the second sample data; A first calculation unit is configured to perform an interaction index prediction score corresponding to the sample data group of each sample cross cluster, determine a target loss value corresponding to each sample cross cluster, and perform a fusion process on the target loss value corresponding to each sample cross cluster to obtain a first loss value; A second calculation unit is configured to determine a second loss value according to a first interaction indicator prediction score corresponding to the first sample data of each sample cross cluster; The model training unit is configured to perform iterative training on the content recommendation model to be trained according to the first loss value and the second loss value to obtain a trained content recommendation model.

11. A content recommendation device, characterized in that: include: A content acquisition unit is configured to acquire the content to be recommended corresponding to the account to be recommended; A score output unit, configured to input the account features of the account to be recommended and the content features of the content to be recommended into a trained content recommendation model for prediction processing, so as to obtain a predicted score of the interaction index of the content to be recommended; the trained content recommendation model is obtained by training the content recommendation model training method according to any one of claims 1 to 8; The content recommendation unit is configured to filter out target content from the content to be recommended according to the interaction indicator prediction score, and recommend the target content to the account to be recommended.

12. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the content recommendation model training method as described in any one of claims 1 to 8, or the content recommendation method as described in claim 9.

13. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the content recommendation model training method as described in any one of claims 1 to 8, or the content recommendation method as described in claim 9.

14. A computer program product, comprising instructions, characterized in that: When the instruction is executed by a processor of an electronic device, the electronic device is enabled to execute the content recommendation model training method as described in any one of claims 1 to 8, or the content recommendation method as described in claim 9.

Images