Recommended model training method, information recommendation method, and related products

Abstract

The application relates to a recommendation model training method, an information recommendation method, equipment, a readable storage medium and a program product, relates to the technical field of artificial intelligence, and can effectively improve information recommendation efficiency. The recommendation model training method comprises the following steps: inputting account attributes of a plurality of resource acquisition users on different clients of different application types and account behaviors triggered by the plurality of resource acquisition users for resource publishing accounts of different clients into a feature extraction module; based on the output of the feature extraction module, obtaining a first similarity of account attribute features of the same resource acquisition users on different clients, a second similarity of account behavior features of the same resource acquisition users on different clients, and a third similarity of account attribute features of different resource acquisition users on different clients; determining a representation enhancement loss according to the first similarity, the second similarity and the third similarity, adjusting model parameters of a recommendation model according to the representation enhancement loss, and obtaining a trained recommendation model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a recommendation model training method, an information recommendation method, a device, a computer equipment, a computer readable storage medium and a computer program product. BACKGROUND

[0002] With the development of computer technology and artificial intelligence technology, recommendation systems are increasingly valued in daily information recommendation. A recommendation system provides accurate and personalized content push for users through relevant analysis. In related technologies, a recommendation system is mainly built around information provided by a single client, and different information corresponding to different business scenarios provided by a single client is used to accurately push relevant content in different scenarios to users.

[0003] However, the recommendation system built in the above manner can only recommend information to users of a single client, and needs to be rebuilt when a new client is used, which cannot achieve efficient information recommendation. SUMMARY

[0004] Therefore, it is necessary to provide a recommendation model training method, an information recommendation method, a computer equipment, a computer readable storage medium and a computer program product to solve the above technical problems.

[0005] In a first aspect, the present application provides a recommendation model training method, comprising:

[0006] inputting account attributes of a plurality of resource acquisition users in different clients of different application types and account behaviors triggered by the plurality of resource acquisition users for resource publishing accounts of the different clients into a feature extraction module of a recommendation model to obtain account attribute features and account behavior features corresponding to each of the resource acquisition users and each of the clients;

[0007] based on the account attribute features and the account behavior features corresponding to each of the resource acquisition users and each of the clients, obtaining a first similarity of account attribute features of the same resource acquisition user in different clients, a second similarity of account behavior features of the same resource acquisition user in different clients, and a third similarity of account attribute features of different resource acquisition users in different clients; wherein the different resource acquisition users are resource acquisition users triggering the same account behavior for resource publishing accounts associated in different clients;

[0008] determining a representation enhancement loss according to the first similarity, the second similarity and the third similarity, adjusting model parameters of the recommendation model according to the representation enhancement loss to obtain a trained recommendation model.

[0009] In one of the embodiments, the determining the enhanced loss based on the first similarity, the second similarity and the third similarity comprises:

[0010] The third similarity is determined according to the weight of the account behaviors triggered by the different resource acquisition users on the different clients to the resource publishing accounts with the association, and the third similarity is adjusted according to the weight to obtain an adjusted third similarity.

[0011] The enhanced loss is determined based on the first similarity, the second similarity and the adjusted third similarity; the enhanced loss is positively correlated with the first similarity, and is negatively correlated with the second similarity and the adjusted third similarity.

[0012] In one of the embodiments, the weight of the third similarity is determined according to the account behaviors triggered by the different resource acquisition users on the different clients to the resource publishing accounts with the association, and the weight of the third similarity is determined according to the weight of the target account behavior.

[0013] If the account behaviors triggered by the different resource acquisition users on the different clients to the resource publishing accounts with the association include multiple types, the behavior priority of each type of the account behaviors is determined.

[0014] The target account behavior is determined according to the behavior priority meeting the priority condition, and the weight of the third similarity is obtained according to the weight corresponding to the target account behavior.

[0015] In one of the embodiments, the account attribute of the multiple resource acquisition users on the different clients comprises the account attribute information of each of the resource acquisition users on the first client and the account attribute information of each of the resource acquisition users on the second client.

[0016] The account behaviors triggered by the multiple resource acquisition users to the resource publishing accounts on the different clients comprise the account behaviors triggered by each of the resource acquisition users to the resource publishing accounts on the first client and the account behaviors triggered by each of the resource acquisition users to the resource publishing accounts on the second client.

[0017] In one of the embodiments, the account attribute features of the same resource acquisition user on the different clients comprise the account attribute features of the first client account and the account attribute features of the second client account corresponding to the same resource acquisition user.

[0018] The account behavior features of the same resource acquisition user on the different clients comprise the account behavior features of the first client account and the account behavior features of the second client account corresponding to the same resource acquisition user.

[0019] The account attribute features of the different resource acquisition users in different clients include: account attribute features of a first client account corresponding to the different resource acquisition users and account attribute features of a second client account.

[0020] In one of the embodiments, the first similarity of the account attribute features of the same resource acquisition user in different clients, the second similarity of the account behavior features of the same resource acquisition user in different clients, and the third similarity of the account attribute features of different resource acquisition users in different clients are determined based on the account attribute features and the account behavior features of each of the resource acquisition users and each of the clients, including:

[0021] determining the first similarity between the account attribute features of the first client account corresponding to the same resource acquisition user and the account attribute features of the second client account;

[0022] determining the second similarity between the account behavior features of the first client account corresponding to the same resource acquisition user and the account behavior features of the second client account;

[0023] determining the third similarity between the account attribute features of the first client account corresponding to the different resource acquisition users and the account attribute features of the second client account.

[0024] In a second aspect, the present application further provides an information recommendation method, including:

[0025] determining a plurality of candidate resource publishing accounts of a first client in response to an account recommendation request for a first client target account of a resource acquisition user; the account recommendation request is generated after a second client target account of the same resource acquisition user triggers a preset event in a second client;

[0026] for each of the candidate resource publishing accounts, obtaining an account information combination corresponding to the candidate resource publishing account according to resource publishing account attribute information of the candidate resource publishing account and account attribute information and account behavior information of the first client target account;

[0027] inputting the account information combination into a trained recommendation model, obtaining a feature combination corresponding to the account information combination by a feature extraction module of the recommendation model, and determining a recommendation index of the candidate resource publishing account according to the feature combination; wherein the recommendation model is trained by the recommendation model training method in any one of claims 1 to 6;

[0028] determining resource publishing account recommendation information provided to the first client target account according to the recommendation index of each of the candidate resource publishing accounts.

[0029] In one embodiment, the account recommendation request is generated and sent by the second client target account after the second client target account of the same resource acquisition user triggers a preset event on the second client.

[0030] And / or,

[0031] The account recommendation request is generated and sent by the second client target account after the second client target account of the same resource acquisition user triggers a preset event on the second client.

[0032] In a third aspect, the present application also provides a computer device, comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of the recommendation model training method or the information recommendation method when executing the computer program.

[0033] In a fourth aspect, the present application also provides a computer readable storage medium, which stores a computer program, and the computer program implements the steps of the recommendation model training method or the information recommendation method when executed by a processor.

[0034] In a fifth aspect, the present application also provides a computer program product, comprising a computer program, and the computer program implements the steps of the recommendation model training method or the information recommendation method when executed by a processor.

[0035] The above recommendation model training method, information recommendation method, computer device, computer readable storage medium and computer program product can obtain the first similarity of the account attribute features of the same resource acquisition user on different clients, the second similarity of the account behavior features of the same resource acquisition user on different clients, and the third similarity of the account attribute features of different resource acquisition users on different clients, and adjust the recommendation model parameters according to the first similarity, the second similarity and the third similarity. The user's similarity in multi-end behavior, the difference between user attributes and the interactivity of user attributes can be described in a contrast learning manner, which helps the model to supplement and strengthen the features of the current client user by using the information of similar users learned from other clients, and to accurately and flexibly push information in multiple clients according to the strengthened features, thereby effectively improving the information recommendation efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present application or the related art. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other related drawings can also be obtained from these drawings without creative labor.

[0037] Figure 1 A flowchart of a recommendation model training method in an embodiment;

[0038] Figure 2 A schematic diagram of the architecture of a recommendation model in an embodiment;

[0039] Figure 3 A flowchart of an information recommendation method in an embodiment;

[0040] Figure 4a A schematic diagram of a music application interface in an embodiment;

[0041] Figure 4b A schematic diagram of a karaoke application interface in an embodiment;

[0042] Figure 4c A schematic diagram of a music application live interface in an embodiment;

[0043] Figure 5 A flowchart of another information recommendation method in an embodiment;

[0044] Figure 6 An internal structure diagram of a computer device in an embodiment;

[0045] Figure 7 An internal structure diagram of another computer device in an embodiment. DETAILED DESCRIPTION

[0046] In order to make the purposes, technical solutions and advantages of the present application clearer, the following will further describe the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0047] It should 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 for analysis, stored data, displayed data, played data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0048] In an embodiment, asFigure 1 As shown, a recommendation model training method is provided, and the embodiment is exemplified by applying the method to a terminal. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction of the terminal and the server. In the embodiment, the method includes the following steps:

[0049] S101, input the account attributes of a plurality of resource acquisition users in different client applications of different application types and the account behaviors triggered by the resource publishing accounts of the plurality of resource acquisition users for different clients to a feature extraction module of a recommendation model to obtain account attribute features and account behavior features corresponding to each resource acquisition user and each client.

[0050] Mobile applications (apps) are deployed on terminals through their clients and can provide diversified services and resources for users and perfect the deficiencies and personalization of original systems. In related technologies, the data in the business scenarios provided by the clients can accurately depict the style of the single client, and the data of other clients is often difficult to align for use, which may face the situation that the users of different clients overlap less and the resources such as services and goods provided by different clients overlap less. Therefore, the traditional recommendation system is mainly constructed around the information provided by a single client, and the corresponding information provided by the single client for different business scenarios is used to accurately push relevant content in different scenarios to users. For example, multi-scenario joint modeling of the same client is based on joint modeling of the same content data of the same client in different scenarios to mine the common and different characteristics between scenarios. This recommendation system construction method conforms to the style of most clients. Of course, in some related technologies, joint modeling of multiple terminal apps is also involved, such as a recommendation system for marketing information or promotion information. However, since the promotion platform strictly limits the information format related to marketing and promotion information to ensure the data format consistency of the promotion information in different clients, the recommendation system construction method is more similar to multi-scenario modeling of a single client, and the feature system and user behavior in different business scenarios are less concerned about the use of scenario information.

[0051] However, the recommendation system constructed in the above manner can only recommend information to the users of a single client or recommend information to multiple clients with basically consistent feature systems or user behaviors. When a user uses a new client of a different application type, if the user behavior or the feature system of the client differs greatly, the recommendation system often needs to be reconstructed due to the small intersection of users.

[0052] To this end, in the embodiments of the present application, a plurality of different clients for joint modeling of multiple terminals can be determined first.

[0053] Different clients can be application clients of different application types, in some examples, different clients can include clients respectively developed by different developers, or clients respectively developed by the same developer for different use purposes, and different clients can be used in the same or different scenarios; the plurality of different clients is two or more, and the plurality of different clients can be mutually trusted.

[0054] Each client can provide an account for a user, for the sake of distinction in the following, the account used to obtain resources in the embodiment is referred to as a resource obtaining account, the user using the resource obtaining account is referred to as a resource obtaining user, the account used to publish or provide resources is referred to as a resource publishing account, and the user using the resource publishing account is referred to as a resource obtaining user; accordingly, the resource obtaining user can obtain the required resources from the resources published by the resource publishing account through the resource obtaining account owned by the user in the client, and the resources can be real-time or pre-generated resources, such as audio live broadcast, video live broadcast, recorded audio, recorded video, image, text, or one or more of the above. It can be understood that the resource obtaining user, the resource publishing user, the resource obtaining account, and the resource publishing account in the embodiment are used to distinguish users and accounts, and in some examples, the types of users and accounts can change, for example, when user A uses his account A to publish resources, at this time, he can be classified as a resource publishing user and a resource publishing account, when user A stops publishing resources and uses account A to browse media resources pushed by the platform, at this time, he can be classified as a resource obtaining user and a resource obtaining account.

[0055] In addition, the plurality of different clients can provide different services or goods, and the behavior types of the account behaviors triggered by the resource obtaining user on the plurality of different clients for the resource publishing account can be the same. For example, for client A and client B, client A is a singing app that can provide various services for real users to interact, such as users entering a video virtual space corresponding to a show video anchor to watch live broadcast; client B is a song listening app that can provide recorded songs published by the resource publishing account, and can also provide an audio virtual space for the resource obtaining user to enter the audio virtual space to listen to live songs, in this scenario, the resources obtained by the user from client A and client B are different: client A provides video live broadcast and user interaction services, and client B provides audio live broadcast and song listening services. However, the resource obtaining user can trigger corresponding account behaviors such as following, browsing, and evaluating one or more account behaviors for the resource publishing account in the client through his account in client A and client B.

[0056] After determining a plurality of different clients for multi-end joint modeling, for each client, a plurality of resource acquisition user account attributes of the resource acquisition user on the client and a plurality of account behaviors triggered by the resource acquisition user on the client for a resource publishing account can be obtained.

[0057] The account attribute is attribute information corresponding to the account of the resource acquisition user on the corresponding client, and the account attribute can represent the characteristics of the user on the corresponding client, such as basic portrait information of the account, including one or more of account identifier, age, gender, and education, and the account attribute can also include statistical features for the account, such as the duration of the account watching, the number of virtual resources (such as virtual props) sent to the resource publishing account, and the like. The account behavior can include one or more of the following: click behavior, viewing behavior, interaction behavior, and payment behavior.

[0058] Further, the account attributes of the plurality of resource acquisition users on different clients and the account behaviors triggered by the plurality of resource acquisition users for the resource publishing accounts of different clients can be input to a feature extraction module of a recommendation model for feature extraction to obtain account attribute features and account behavior features corresponding to each resource acquisition user and each client.

[0059] S102, based on the account attribute features and the account behavior features corresponding to each resource acquisition user and each client, a first similarity of the account attribute features of the same resource acquisition user on different clients, a second similarity of the account behavior features of the same resource acquisition user on different clients, and a third similarity of the account attribute features of different resource acquisition users on different clients are obtained; wherein the different resource acquisition users are resource acquisition users who trigger the same account behavior on the associated resource publishing accounts of different clients.

[0060] In some examples, the associated resource publishing account can be a resource publishing account owned by the same resource publishing user on different clients.

[0061] After obtaining the account attribute features and the account behavior features corresponding to each resource acquisition user and each client, multi-end app joint modeling can be performed based on contrast learning. Contrast learning uses similarity and dissimilarity to enable the model to map similar instances closely together in the latent feature space, while distinguishing different or irrelevant instances. The embodiments of the present application use contrast learning to strengthen the influence of overlapping user behaviors, including the similarity, difference, and interactivity of user behaviors.

[0062] Specifically, although different multiple apps have inconsistent feature systems, large differences in user behaviors, and few intersection users, the inventors have found in practice that the same user has similarities in multi-end behaviors, the user attributes of different clients have differences, and the user attributes have intercommunication. The similarities of the same user refer to the same user having similar behaviors on different clients, which can be described by the similar situations of the account behaviors of the same user on different clients. The differences in user attributes are introduced by the user on different clients, which are caused by the differences in the essential attributes of different clients, and naturally exist differences in the information carried by the user attributes. The intercommunication of user attributes can be understood as different users triggering the same behavior for associated resource publishing users, and it can be considered that the user attributes of these different users have similarities, that is, due to the similarity of the user attributes of different users, the same account behavior is performed on the account of the same resource publishing user in different clients. In some embodiments, the similarity of the user can be described by the behavior type of the same account behavior.

[0063] Based on the above concept, the embodiment of the present application can obtain the similarity of the account attribute features of the same resource obtaining user on different clients on the one hand, for convenience of distinction, the similarity is called the first similarity. On the other hand, the similarity of the account behavior features of the same resource obtaining user on different clients can be obtained, for convenience of distinction, the similarity is also called the second similarity. On the other hand, a specific different resource obtaining user can also be determined, and then the similarity of the account attribute features of the different resource obtaining user on different clients can be obtained, for convenience of distinction, the similarity is called the third similarity. Among them, the different resource obtaining user in the embodiment is a resource obtaining user who triggers the same account behavior for the resource publishing account associated with different clients.

[0064] S103, determining an enhanced loss according to the first similarity, the second similarity and the third similarity, adjusting the model parameters of the recommendation model according to the enhanced loss, and obtaining a trained recommendation model.

[0065] Specifically, the enhanced loss can be determined according to the first similarity, the second similarity and the third similarity , and the model parameters of the recommendation model are adjusted according to the enhanced loss and the back propagation algorithm, and a trained recommendation model is obtained.

[0066] In practical applications, the recommendation model can be a multi-objective model, which is used to determine the recommendation index of the resource publishing account to the resource obtaining user when triggering the specified account behavior is expected according to the features extracted from the input information, for example Figure 2In the recommended model shown, a first recommended module configured for a first client and a second recommended module configured for a second client are included, and the first recommended module and the second recommended module can both output a recommended index for a resource publishing account when a resource acquisition user triggers a click behavior, a viewing behavior, an interaction behavior, and a payment behavior on the resource publishing account. The structure of the recommended module (the first recommended module and / or the second recommended module) can be flexibly set according to actual conditions. For example, the recommended module can include a scene customized gate control network layer (CGC) and a cross layer. The scene CGC is a multi-target processing method according to an end scene, and scene expert modules and public expert modules are allocated according to different scenes. The cross layer is used to process input features based on a feature cross method, such as a cross method combined with a factorized machine (FM). In some examples, a representation layer can be provided in the recommended model to determine the first similarity, the second similarity, and the third similarity.

[0067] In some example embodiments, when training the recommended model, a recommended index label can also be introduced. According to the difference between the recommended index output by the first recommended module and the second recommended module and the recommended index label, a recommended module loss is determined, a model loss value is determined by combining the recommended module loss and the representation enhancement loss, and the model parameters of the recommended model are adjusted.

[0068] In the recommended model training method described above, by obtaining the first similarity of account attribute features of the same resource acquisition user on different clients, the second similarity of account behavior features of the same resource acquisition user on different clients, and the third similarity of account attribute features of different resource acquisition users on different clients, and adjusting the recommended model parameters according to the first similarity, the second similarity, and the third similarity, the similarity of user behaviors on multiple terminals, the difference of user attributes, and the interactivity of user attributes can be described through a contrastive learning method. This helps the model to supplement and strengthen the features of the current client user by using the information of similar users learned from other clients, and to accurately and flexibly push information on multiple clients according to the strengthened features, thereby effectively improving the information recommendation efficiency.

[0069] In one example embodiment, in step S103, determining the representation enhancement loss according to the first similarity, the second similarity, and the third similarity can include the following steps:

[0070] S201, determining a weight of the third similarity according to the account behaviors of different resource acquisition users triggered on different clients on the resource publishing accounts associated with each other, and adjusting the third similarity according to the weight to obtain an adjusted third similarity.

[0071] In this step, different resource acquisition users can be determined, i.e., resource acquisition users triggering account behaviors of the associated resource publishing account in different clients. For example, for a resource publishing user U, the user opens an anchor account in client A, client B, and client C, which are anchor account 1, anchor account 2, and anchor account 3, respectively. Different resource acquisition users trigger account behaviors of anchor account 1, anchor account 2, and anchor account 3 through their client accounts in the corresponding clients. Resource acquisition user 1 triggers a follow behavior for anchor account 1, resource acquisition user 2 triggers a payment behavior for anchor account 2, and resource acquisition user 3 triggers a follow behavior for anchor account 3. Resource acquisition user 1 and resource acquisition user 3 can be determined as different resource acquisition users, and the same account behavior triggered is determined as a follow behavior.

[0072] In actual applications, the degree of description of similar behaviors differs for different account behaviors. Specifically, for simple behaviors or behaviors not concerned in complex business scenarios, such as interactive behaviors or payment behaviors, the behavior similarity of different users can be more accurately constructed. For complex behaviors or behaviors not concerned in complex business scenarios, the accuracy of constructing the behavior similarity can be relatively reduced. For example, specific browsing behaviors or click behaviors, because the browsing time and the number of clicks of different users are difficult to accurately approach, the weight of the behaviors can also be correspondingly reduced.

[0073] Based on this, after determining the account behaviors triggered by different resource acquisition users in different clients for the associated resource publishing account, the weight of the third similarity can be determined according to the account behavior. In some examples, the weight of the third similarity can be positively correlated with the level of description accuracy of the account behavior, that is, the higher the level of description accuracy of the account behavior, the greater the weight. Then, the third similarity is adjusted according to the weight to obtain an adjusted third similarity. For example, the product of the third similarity and the weight of the third similarity can be obtained, and the product is taken as the adjusted third similarity.

[0074] S202, determining an enhanced loss according to the first similarity, the second similarity, and the adjusted third similarity; the enhanced loss is positively correlated with the first similarity, and negatively correlated with the second similarity and the adjusted third similarity.

[0075] In this step, after obtaining the first similarity, the second similarity, and the adjusted third similarity, the enhanced loss can be determined. In some exemplary embodiments, taking joint modeling of two clients as an example, the enhanced loss is determined as follows: The enhanced loss can be as follows:

[0076]

[0077] in, The first similarity score, For the second similarity, The third similarity The weight of the third similarity, User U, who obtains resources, is on the first client. Account attribute characteristics, User U obtains resources on the second client. Account attribute characteristics, User U, who obtains resources, is on the first client. Account behavior characteristics User U obtains resources on the second client. Account behavior characteristics User U' obtains resources on the first client Account attribute characteristics, User U'' in the second client is used to obtain resources. The account attribute characteristics in the data are as follows: resource acquisition user U'' and resource acquisition user U'' are resource acquisition users who have triggered the same account behavior on the resource publishing account with the same relationship on the second client and the first client respectively. Resource acquisition user U' and resource acquisition user U can be the same user or different users. Resource acquisition user U'' and resource acquisition user U are different users.

[0078] In this embodiment, on the one hand, by making the representation enhancement loss positively correlated with the first similarity and negatively correlated with the second and third similarities, the recommendation model can make the account behavior features output by the same user on different clients more similar, and make the account attribute features of different users with the same attributes on different clients more similar. This facilitates the close mapping of the behavior features of the same user or the attribute features of similar users together, enabling more accurate customer segmentation based on user characteristics and behavior. On the other hand, it also makes the account attribute features output by the recommendation model for the same user on different clients more distinct, increasing the discriminative power of account attribute features across different clients. This makes it easier for the recommendation model to recommend information that matches the characteristics of a specific client by combining the attributes of that client. Furthermore, by adjusting the weights differently to adjust the impact of different account behaviors on the loss, it is helpful to characterize the similarity of various user behaviors with different levels of precision.

[0079] In an exemplary embodiment, in step S201, determining the weight of the third similarity based on the behavior type of accounts of the same type may include the following steps:

[0080] If the account behaviors triggered by the resource publishing account associated with the different resource acquisition users in different clients include multiple types, the behavior priority of each type of account behavior is determined; a target account behavior that meets the priority condition is determined according to the behavior priority, and the weight of the third similarity is obtained according to the weight corresponding to the target account behavior.

[0081] In actual application, the resource acquisition users can trigger multiple same account behaviors for the resource publishing account associated with the different resource acquisition users. For example, in the above example, the resource acquisition user 1 triggers the attention behavior and the watching behavior for the anchor account 1, and the resource acquisition user 3 also triggers the attention behavior and the watching behavior for the anchor account 3. In the case that the triggered account behaviors include multiple types, the behavior priority of each type of account behavior can be determined, and then the target account behavior that meets the priority condition is determined, and then the weight of the third similarity can be determined according to the weight of the target account behavior. For example, the behavior priorities of four types of account behaviors, i.e., pay, interaction, time and click, are as follows:

[0082]

[0083] When the interaction behavior and the click behavior exist at the same time, the interaction behavior with the highest behavior priority can be determined as the target account behavior, and then the corresponding weight is obtained.

[0084] In this embodiment, the weight of the third similarity is obtained according to the weight corresponding to the target account behavior that meets the priority condition according to the behavior priority, so that the third similarity can be adjusted differently according to the priority when multiple same types of account behaviors exist, and the matching of the third similarity with the actual scene is improved.

[0085] In an exemplary embodiment, the account attributes of the multiple resource acquisition users in different clients include: the account attribute information of each resource acquisition user in the multiple resource acquisition users in the first client, and the account attribute information of each resource acquisition user in the second client; the account behaviors triggered by the resource publishing account associated with the multiple resource acquisition users in different clients include: the account behaviors triggered by each resource acquisition user in the multiple resource acquisition users for the resource publishing account in the first client, and the account behaviors triggered by each resource acquisition user for the resource publishing account in the second client.

[0086] In a specific implementation, a resource acquisition user can be determined, and the resource acquisition user holds a corresponding resource acquisition account on the first client and the second client. To distinguish the resource acquisition accounts on different clients, the resource acquisition account on the first client is referred to as a first client account, and the resource acquisition account on the second client is referred to as a second client account. It can be understood that, since the first client and the second client are different clients, the first client account and the second client account correspond to the same resource acquisition user, but the account attributes and account behaviors corresponding to each account can be different.

[0087] Then, for the resource acquisition user, relevant information of the resource acquisition user on different clients can be acquired respectively. Specifically, the account attribute information of the resource acquisition user on the first client account and the account behavior triggered by the first client account for the resource publishing account of the first client can be acquired, and the account attribute information of the resource acquisition user on the second client account and the account behavior triggered by the second client account for the resource publishing account of the second client can be acquired. Then, the above information corresponding to each of the plurality of resource acquisition users can be input to a feature extraction module of a recommendation model for feature extraction.

[0088] Taking the resource publishing account as an anchor account as an example, referring to Figure 2 For each resource acquisition user, the account identifier of the resource acquisition user on the client A and the client B, the account identifier of the anchor account triggering the account behavior, and the account behavior in the live room in different clients can be concatenated together as sample data of the resource acquisition user, and input to an embedding vector layer (i.e., a feature extraction module) of a resource acquisition model for feature extraction.

[0089] In this embodiment, by acquiring the account attribute information and the account behavior triggered by the resource publishing account of the resource acquisition user in the first client and the second client, the associated data of the same user in the two clients can be acquired, thereby providing a basis for cross-application multi-end modeling of the first client and the second client.

[0090] Correspondingly, in some embodiments, the account attribute features of the same resource acquisition user on different clients include: the account attribute features of the first client account and the second client account corresponding to the same resource acquisition user; the account behavior features of the same resource acquisition user on different clients include: the account behavior features of the first client account and the second client account corresponding to the same resource acquisition user; and the account attribute features of different resource acquisition users on different clients include: the account attribute features of the first client account and the second client account corresponding to different resource acquisition users.

[0091] In a specific implementation, after obtaining the account attribute features of each first client account and the account attribute features of each second client account output by the feature extraction module, the account attribute features of the first client account and the account attribute features of the second client account corresponding to the same resource acquisition user can be determined as the account attribute features of the same resource acquisition user on different clients through device information or other identification information.

[0092] On the other hand, the account behavior features of the first client account and the account behavior features of the second client account corresponding to the same resource acquisition user can also be determined as the account behavior features of the same resource acquisition user on different clients through device information or other identification information.

[0093] On the other hand, the account behavior features of the first client account and the account behavior features of the second client account corresponding to the same resource acquisition user can also be determined as the account behavior features of the same resource acquisition user on different clients through device information or other identification information.

[0094] Therefore, the account attribute features and the account behavior features obtained above can provide a basis for constructing the user behavior similarity, the user attribute difference, and the user attribute interactivity between two clients, and the implementation is simple and convenient, which helps to improve the construction efficiency of the recommendation model.

[0095] Correspondingly, step S102 can include the following steps:

[0096] determining a first similarity between the account attribute features of the first client account and the account attribute features of the second client account corresponding to the same resource acquisition user; determining a second similarity between the account behavior features of the first client account and the account behavior features of the second client account corresponding to the same resource acquisition user; and determining a third similarity between the account attribute features of the first client account and the account attribute features of the second client account corresponding to different resource acquisition users.

[0097] Specifically, for the difference of user attributes of the same user on different clients, a feature similarity between the account attribute features of the first client account and the second client account can be determined, and the similarity is taken as a first similarity. For the similarity of user behaviors on multiple terminals, a feature similarity between the account behavior features of the first client account and the second client account corresponding to the same resource obtaining user can be determined, and the feature similarity is taken as a second similarity. For the intercommunication of different user attributes, a feature similarity between the account attribute features of the first client account and the second client account corresponding to different resource obtaining users can be determined, and the feature similarity is taken as a third similarity. In this way, the account behavior features output for the same user in the first client and the second client and the account attribute features of similar users on the first client and the second client can be as close as possible, so that the user information of the current client can be supplemented and strengthened by using the user information of other clients, while the account attribute features of the same user on different clients are far away, the discrimination of the account attribute features under a specific client is increased, and the accuracy of subsequent information recommendation is improved.

[0098] In one embodiment, as shown in Figure 3 A method for information recommendation is also provided. In this embodiment, the method is applied to a server. It should be understood that the method can also be applied to a terminal and a system including a terminal and a server, and is realized through the interaction of the terminal and the server. In this embodiment, the method includes the following steps:

[0099] S301, in response to an account recommendation request for a first client target account of a resource obtaining user, determining a plurality of candidate resource publishing accounts of the first client; the account recommendation request is generated after a second client target account of the same resource obtaining user in a second client triggers a preset event.

[0100] In this embodiment, the first client target account is an account of the first client, the second client target account is an account of the second client, and the first client target account and the second client target account correspond to the same resource obtaining user; the first client and the second client can be clients that provide training account data when a recommendation model is trained.

[0101] In a specific implementation, the terminal is deployed with the first client and the second client, and the resource obtaining user can perform relevant account behaviors on the first client through the first client target account owned by the resource obtaining user, or perform relevant account behaviors on the second client through the second client target account.

[0102] When the second client detects that the second client target account triggers the preset event, a corresponding account recommendation request can be obtained, which can request to obtain a resource publishing account related to the resource for the user owning the first client target account.

[0103] In some embodiments, the account recommendation request can be generated and sent by the first client triggered by the second client target account after the second client target account of the same resource acquisition user triggers the preset event in the second client.

[0104] In a specific implementation, the first client and the second client respectively communicate with the corresponding server, and the communication is independent of each other. For example, the first client communicates with the server corresponding to the first client, and the second client does not communicate with the server corresponding to the first client. At this time, the second client generates an account recommendation request by triggering the first client, which can avoid other clients communicating with the server that does not correspond to the server, and realize cross-client information pushing while ensuring the communication security and data security of the client and the server. For example, when the "music application 1" detects that the user clicks the button "open" in Figure 4a , it can be determined that the second client target account triggers the preset event, and the "K-song application 2" sends an account recommendation request. For another example, when the "K-song application 2" detects that the user clicks the button 401 "get" in Figure 4b , it can be determined that the second client target account triggers the preset event, and the "music application 1" sends an account recommendation request. According to the information recommendation method provided in the embodiments of the present application, the resource publishing account of the "music application 1" is recommended and jumps to the corresponding live room (as shown in Figure 4c ).

[0105] Of course, the account recommendation request can also be generated and sent directly by the second client target account after the second client target account of the same resource acquisition user triggers the preset event in the second client. For example, when the server and the second client trust each other, the second client target account can generate and send the account recommendation request to the server directly. In this way, the flexibility of the second client to recommend the first client resource publishing account can be improved.

[0106] S302, for each candidate resource publishing account, according to the resource publishing account attribute information of the candidate resource publishing account, and the account attribute information and the account behavior information of the first client target account, obtain the account information combination corresponding to the candidate resource publishing account.

[0107] After obtaining a plurality of candidate resource publishing accounts, for each resource publishing account, the resource publishing account attribute information of the candidate resource publishing account, and the account attribute information and the account behavior information of the target account can be obtained, and the account information combination corresponding to the candidate resource publishing account can be obtained. For example, as shown inFigure 5 As shown, the first client requests a corresponding online anchor list for the first client target account, combines the first client target account, the client (i.e., the first client) and each anchor into a request pair, and then queries the corresponding client feature platform to obtain account attribute information, account behavior information (such as a follow sequence, a comment sequence and a payment sequence), client information and anchor account attribute information of the first client target account, and generates an account information combination (i.e., a feature pair in Figure 5 ).

[0108] In S303, the account information combination is input into the trained recommendation model, a feature combination corresponding to the account information combination is obtained by a feature extraction module of the recommendation model, and a recommendation index of a candidate resource publishing account is determined according to the feature combination.

[0109] The recommendation model can be trained by the recommendation model training method in one or more embodiments described above.

[0110] After obtaining the account information combination corresponding to each of the plurality of candidate resource publishing accounts, the account information combination can be input into the trained recommendation model, the account information combination is feature-extracted by a feature extraction module of the recommendation model, a feature combination corresponding to the account information combination is obtained, and a recommendation index of a candidate resource publishing account is determined according to the feature combination.

[0111] In S304, resource publishing account recommendation information provided to the first client target account is determined according to the recommendation index of each candidate resource publishing account.

[0112] After obtaining the recommendation indexes of the plurality of candidate resource publishing accounts, resource publishing account recommendation information provided to the first client target account can be obtained according to the plurality of recommendation indexes. For example, in some embodiments, the recommendation model can score multiple targets in different scenarios at the same time, and in actual application, the scores of the multiple targets can be automatically obtained according to the scene, then the scores of the multiple targets are weighted and fused to obtain a recommendation index, and then the fused scores are sorted from high to low, and then pushed to the user in this order for display.

[0113] The above information recommendation method determines a plurality of candidate resource publishing account numbers of the first client in response to an account recommendation request for a first client target account number of a resource obtaining user, wherein the account recommendation request is generated by a second client target account number of a second client of the same resource obtaining user after triggering a preset event; then, for each candidate resource publishing account number, the account information combination corresponding to the candidate resource publishing account number can be obtained according to the resource publishing account number attribute information of the candidate resource publishing account number, and the account attribute information and the account behavior information of the first client target account number; then, the account information combination is input into the trained recommendation model, the feature extraction module of the recommendation model obtains the feature combination corresponding to the account information combination, and determines the recommendation index of the candidate resource publishing account number according to the feature combination; then, the resource publishing account recommendation information provided to the first client target account number can be determined according to the recommendation indexes of the candidate resource publishing account numbers. In the embodiment, the similarity of the user behavior in multiple terminals, the difference of the user attributes, and the intercommunication of the user attributes can be constructed through the recommendation model. When the account information combination of the user in the first client is obtained, the feature information of the current account can be supplemented and strengthened by using the information of similar or same users learned from other terminals through the recommendation model, so that the information recommendation can be more accurately performed according to the strengthened features, and the accuracy and efficiency of the information recommendation can be improved.

[0114] The inventors have found in practice that the recommendation by the recommendation model provided in the embodiments of the present application can effectively optimize the information recommendation effect for different users, and users can more quickly obtain resource information matching their expected preferences and habits.

[0115] It should be understood that, although each step in the flowchart involved in each embodiment as described above is displayed in sequence according to the arrow, these steps are not necessarily executed in sequence according to the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other sequences. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least part of other steps or steps or stages in other steps.

[0116] In an exemplary embodiment, a computer device, which can be a server, is provided, and an internal structure diagram of the computer device can be as shown in Figure 6As shown in the figure. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through the system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control ability. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the account data of a plurality of clients. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement a recommendation model training method or an information recommendation method.

[0117] In an exemplary embodiment, a computer device, which can be a terminal, is provided, and an internal structure diagram thereof can be as shown in the figure. Figure 7 As shown in the figure. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. Among them, the processor, the memory and the input / output interface are connected through the system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control ability. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner, and the wireless manner can be realized through WIFI, mobile cellular network, near field communication (Near Field Communication, NFC) or other technologies. The computer program is executed by the processor to implement a recommendation model training method or an information recommendation method. The display unit of the computer device is used to form a visually visible picture, which can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0118] Those skilled in the art can understand that, Figure 6 and Figure 7The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0119] In an embodiment, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps in the above method embodiments when executing the computer program.

[0120] In an embodiment, a computer readable storage medium is provided, storing a computer program, and the computer program implementing the steps in the above method embodiments when executed by a processor.

[0121] In an embodiment, a computer program product is provided, comprising a computer program, and the computer program implementing the steps in the above method embodiments when executed by a processor.

[0122] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiment methods. Any reference to memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. The non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. The volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., without being limited thereto.

[0123] The technical features of the above embodiments can be combined arbitrarily. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present application.

[0124] The above-described embodiments are merely illustrative of several embodiments of the present application, and the description is relatively specific and detailed, but should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A method for training a recommendation model, characterized in that, The method comprises: inputting account attributes of a plurality of resource acquisition users in different clients of different application types and account behaviors triggered by the resource acquisition users for resource publishing accounts of the different clients into a feature extraction module of a recommendation model to obtain account attribute features and account behavior features corresponding to each of the resource acquisition users and each of the clients; based on the account attribute features and the account behavior features corresponding to each of the resource acquisition users and each of the clients, obtaining a first similarity of account attribute features of the same resource acquisition user in different clients, a second similarity of account behavior features of the same resource acquisition user in different clients, and a third similarity of account attribute features of different resource acquisition users in different clients; wherein the different resource acquisition users are resource acquisition users triggering the same account behavior for resource publishing accounts associated with different clients; determining a representation enhancement loss according to the first similarity, the second similarity, and the third similarity, and adjusting model parameters of the recommendation model according to the representation enhancement loss to obtain a trained recommendation model.

2. The method of claim 1, wherein, The determination of the representation enhancement loss according to the first similarity, the second similarity, and the third similarity comprises: determining a weight of the third similarity according to account behaviors triggered by the different resource acquisition users for resource publishing accounts associated with different clients, and obtaining an adjusted third similarity according to the weight of the third similarity; determining a representation enhancement loss according to the first similarity, the second similarity, and the adjusted third similarity; the representation enhancement loss is positively correlated with the first similarity and negatively correlated with the second similarity and the adjusted third similarity.

3. The method of claim 2, wherein, The determination of the weight of the third similarity according to the account behaviors triggered by the different resource acquisition users for resource publishing accounts associated with different clients comprises: if the account behaviors triggered by the different resource acquisition users for resource publishing accounts associated with different clients include multiple types, determining a behavior priority of each type of the account behaviors; determining a target account behavior that satisfies a priority condition according to the behavior priority, and obtaining the weight of the third similarity according to a weight corresponding to the target account behavior.

4. The method of claim 1, wherein, The account attributes of the plurality of resource acquisition users in different clients comprise account attribute information of each of the resource acquisition users in a first client and account attribute information of each of the resource acquisition users in a second client; The account behaviors triggered by the plurality of resource acquisition users for resource publishing accounts of the different clients comprise account behaviors triggered by each of the resource acquisition users for resource publishing accounts of the first client and account behaviors triggered by each of the resource acquisition users for resource publishing accounts of the second client.

5. The method of claim 4, wherein, The account attribute features of the same resource acquisition user in different clients comprise account attribute features of a first client account of the same resource acquisition user and account attribute features of a second client account of the same resource acquisition user. The account behavior features of the same resource acquisition user in different clients include: account behavior features of a first client account corresponding to the same resource acquisition user and account behavior features of a second client account; The account attribute features of the different resource acquisition users in different clients include: account attribute features of a first client account corresponding to the different resource acquisition users and account attribute features of a second client account.

6. The method of claim 5, wherein, The method comprises: The method comprises: The account recommendation request is generated and sent by the second client target account after a preset event triggered by a second client target account of the same resource acquisition user in a second client; The account recommendation request is generated and sent by the second client target account after a preset event triggered by a second client target account of the same resource acquisition user in a second client.

7. An information recommendation method characterized by comprising: The processor executes the computer program to realize the steps of the recommendation model training method of any one of claims 1 to 6 or the steps of the information recommendation method of any one of claims 7 to 8. ​ ​ ​ ​ 8. The method of claim 7, wherein, ​ ​ ​ 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. ​ 10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by a processor to implement the steps of the recommendation model training method trained by any one of claims 1 to 6 or the steps of the information recommendation method of any one of claims 7 to 8.

11. A computer program product comprising a computer program, characterized in that, The computer program is executed by a processor to implement the steps of the recommendation model training method trained by any one of claims 1 to 6 or the steps of the information recommendation method of any one of claims 7 to 8.

Images