Article recommendation method and device, recommendation model training method and device and storage medium

Abstract

The invention provides an article recommendation method and device, a recommendation model training method and device, a storage medium and a program product, can be applied to the technical field of artificial intelligence, and particularly relates to the technical field of multi-field recommendation. The method comprises the following steps: in response to an operation of a target user in a target application, obtaining target user information of the target user and article information of at least one article to be recommended in the target application; performing preference decoupling processing on the article information of the at least one to-be-recommended article to obtain a target domain sharing feature and a target domain specific feature of each to-be-recommended article; determining a target domain enhancement feature corresponding to a target domain according to the target domain to which the target application belongs; according to the target user information and the target domain sharing feature, the target domain specific feature and the target domain enhancement feature of each to-be-recommended article, predicting the recommendation probability of each recommended article; and according to the recommendation probability of the at least one to-be-recommended article, generating an article recommendation list for the target user.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, specifically to the field of multi-field recommendation technology, and more specifically to an item recommendation method, a training method, device, storage medium and program product of a recommendation model. Background Art

[0002] Recommendation systems are widely used in various real-world scenarios. For example, a variety of recommendation functions are applied in different fields such as e-commerce, video, and mini-programs. User behaviors in different recommendation fields can be shared and associated to form a multi-field recommendation function that integrates multiple fields.

[0003] However, a common method for multi-domain recommendation in related technologies is to directly combine user behaviors in multiple fields as input features, so that both favorable and unfavorable potential factors are directly transferred during multi-domain recommendation, thus affecting the accuracy of multi-domain recommendation. Summary of the invention

[0004] In view of this, the present disclosure provides an item recommendation method, a training method, a device, a storage medium, and a program product for a recommendation model.

[0005] One aspect of the present disclosure provides an item recommendation method, comprising: in response to an operation of a target user in a target application, obtaining target user information of a target user and item information of at least one item to be recommended in the target application; performing preference decoupling processing on the item information of at least one item to be recommended to obtain target domain shared features and target domain specific features of each item to be recommended; determining target domain enhanced features corresponding to the target domain according to the target domain to which the target application belongs; predicting the recommendation probability of each recommended item according to the target user information, the target domain shared features, the target domain specific features and the target domain enhanced features of each item to be recommended; and generating an item recommendation list for the target user according to the recommendation probability of at least one item to be recommended.

[0006] According to an embodiment of the present disclosure, preference decoupling processing is performed on the item information of at least one item to be recommended to obtain the target domain shared features and target domain specific features of each item to be recommended, including: performing domain sharing encoding processing on the item information of at least one item to be recommended to obtain the target domain shared features of each recommended item; performing domain specific encoding processing on the item information of at least one item to be recommended to obtain the target domain specific features of each recommended item.

[0007] According to an embodiment of the present disclosure, determining the target domain enhancement features corresponding to the target domain according to the target domain to which the target application belongs includes: determining the target domain enhancement features for the target domain from a domain-feature relationship graph according to the target domain to which the target application belongs, wherein the domain-feature relationship graph includes at least one domain and the domain enhancement features corresponding to each domain.

[0008] According to an embodiment of the present disclosure, predicting the recommendation probability of each recommended item based on target user information, target domain shared features, target domain specific features and target domain enhanced features of each item to be recommended includes: for each item to be recommended, determining the target fusion features of each item to be recommended based on the target domain shared features, target domain specific features and target domain enhanced features; predicting the recommendation probability of each recommended item based on the target user information and the target fusion features of each item to be recommended.

[0009] According to an embodiment of the present disclosure, the above-mentioned item recommendation method also includes: performing semantic feature extraction on the target user information and the item information of at least one item to be recommended to obtain the target user feature and the item feature to be recommended of each item to be recommended; wherein, performing preference decoupling processing on the item information of at least one item to be recommended to obtain the target domain shared feature and the target domain specific feature of each item to be recommended includes: performing preference decoupling processing on the item feature to be recommended of each item to be recommended to obtain the target domain shared feature and the target domain specific feature of each item to be recommended; predicting the recommendation probability of each recommended item according to the target user information, the target domain shared feature, the target domain specific feature and the target domain enhanced feature of each item to be recommended includes: predicting the recommendation probability of each recommended item according to the target user feature, the target domain shared feature, the target domain specific feature and the target domain enhanced feature of each item to be recommended.

[0010] Another aspect of the present disclosure provides a training method for a recommendation model, including: performing preference decoupling processing on semantic features of sample items in respective M sample domains to obtain sample domain shared features and sample domain specific features for each sample item, wherein the sample domain shared features represent features shared across domains, and the sample domain specific features represent features specific to a domain, and M≥2; performing feature fusion processing on the sample domain shared features and sample domain specific features of each sample item to obtain sample fusion features for each sample item; predicting the recommendation probability of each sample item based on sample user features and the sample fusion features of each sample item, wherein the sample user features represent features of users who interact with sample items in the M sample domains; and determining recommendation loss information of the recommendation model based on the recommendation probability of each sample item, iteratively adjusting the network parameters of the recommendation model based on the recommendation loss information until the recommendation loss information satisfies a recommendation iteration stop condition, and using the network parameters obtained when the recommendation iteration stop condition is satisfied as the trained recommendation model.

[0011] According to an embodiment of the present disclosure, the recommendation model includes a preference decoupling module, which includes a domain-shared encoder and a domain-specific encoder corresponding to each of the M sample domains; wherein, preference decoupling processing is performed on the semantic features of the sample items of each of the M sample domains to obtain the sample domain-shared features and sample domain-specific features of each sample item, including: inputting the semantic features of the sample items of each of the M sample domains into the domain-shared encoder, and outputting the sample domain-shared features of each sample item; inputting the semantic features of the sample items of each of the M sample domains into the corresponding domain-specific encoders, and outputting the sample domain-specific features of each sample item.

[0012] According to an embodiment of the present disclosure, the training method of the above-mentioned recommendation model also includes: determining the similarity of specific features between every two sample domains using a first preset formula based on the sample domain specific features of each sample item; adjusting the network parameters of the domain-specific encoder using the specific feature similarity; determining the similarity of shared features between every two sample domains using a second preset formula based on the sample domain shared features of each sample item; and adjusting the network parameters of the domain-shared encoder using the shared feature similarity.

[0013] According to an embodiment of the present disclosure, the recommendation model includes a preference fusion module; feature fusion processing is performed according to the sample domain shared features and sample domain specific features of each sample item to obtain the sample fusion features of each sample item, including: inputting the sample domain shared features and sample domain specific features of each sample item into the preference fusion module, and determining the sample domain enhanced features of each sample item according to the sample domain shared features of each sample item; feature fusion processing is performed according to the sample domain shared features, sample domain specific features and sample domain enhanced features of each sample item, and the sample fusion features of each sample item are output.

[0014] According to an embodiment of the present disclosure, determining the sample domain enhanced features of each sample field according to the sample domain shared features of each sample item includes: fusing the sample domain shared features of each sample item to obtain the sample domain enhanced features; or clustering the sample domain shared features of each sample item to obtain at least one cluster cluster; generating a cluster-level cross-domain graph according to the cluster features of each cluster cluster, wherein the vertices of the cluster-level cross-domain graph correspond to the cluster clusters and the edges represent the similarities between clusters; determining the sample domain enhanced features of each sample field according to the cluster-level cross-domain graph and each cluster cluster.

[0015] According to an embodiment of the present disclosure, determining the sample domain enhanced features of each sample domain according to the cluster-level cross-domain graph and each clustering cluster includes: performing domain sharing preference aggregation processing on the cluster-level cross-domain graph to obtain enhanced cluster features of each clustering cluster, wherein the enhanced cluster feature characterization adds inter-domain shared features to the cluster features; determining the sample domain enhanced features of each sample domain according to the cluster features and enhanced cluster features of each clustering cluster.

[0016] According to an embodiment of the present disclosure, determining the sample domain enhanced features of each sample field according to the cluster features and enhanced cluster features of each cluster cluster includes: for the jth sample field among the M sample fields, determining the jth cluster allocation matrix of the jth sample field according to the cluster features of each cluster cluster, wherein the jth cluster allocation matrix includes the allocation coefficient of each cluster cluster in the jth sample field, 1≤j≤M; determining the jth sample domain enhanced features of the jth sample field according to the j cluster allocation matrices and the enhanced cluster features of each cluster cluster.

[0017] According to an embodiment of the present disclosure, determining the j-th sample domain enhanced feature of the j-th sample domain based on j cluster allocation matrices and the enhanced cluster features of each cluster cluster includes: for each cluster cluster, multiplying the enhanced cluster feature of the cluster cluster with the allocation coefficient of the cluster cluster in the j-th sample domain to obtain the allocation feature of the cluster cluster; and summing the allocation features of each cluster cluster to obtain the j-th sample domain enhanced feature.

[0018] According to an embodiment of the present disclosure, the training method of the recommendation model further includes: determining the sample item enhanced features of each sample item according to the sample domain enhanced features of each sample domain; determining the enhanced domain shared features of the sample item according to the sample domain shared features and the sample item enhanced features of each sample item; for the i-th sample domain among the M sample domains, determining a first preference score for the sample domain specific features according to the sample user features and the sample domain specific features of the sample items in the i-th sample domain; determining a second preference score for the enhanced domain shared features according to the sample user features and the enhanced domain shared features of the sample items; determining fusion loss information for the preference fusion module based on the first preference score and the second preference score, and iteratively adjusting the network parameters of the preference fusion module according to the fusion loss information until the fusion loss information meets the fusion iteration stop condition.

[0019] According to an embodiment of the present disclosure, the recommendation model also includes a semantic extraction module; the above method also includes: obtaining text data of sample items in M ​​sample fields and sample user information of users who interact with the sample items in the M sample fields; inputting the text data of the sample items and the sample user information into the semantic extraction module, and outputting sample user features and semantic features of the sample items in each of the M sample fields.

[0020] Another aspect of the present disclosure provides an item recommendation device, including: a first acquisition module, used to obtain target user information of a target user and item information of at least one item to be recommended in a target application in response to an operation of a target user in a target application; a decoupling processing module, used to perform preference decoupling processing on the item information of at least one item to be recommended, and obtain target domain shared features and target domain specific features of each item to be recommended; a first determination module, used to determine target domain enhanced features corresponding to the target domain according to the target domain to which the target application belongs; a first prediction module, used to predict the recommendation probability of each recommended item according to the target user information, the target domain shared features, the target domain specific features and the target domain enhanced features of each item to be recommended; and a generation module, used to generate an item recommendation list for the target user according to the recommendation probability of at least one item to be recommended.

[0021] Another aspect of the present disclosure provides a training device for a recommendation model, including: a preference decoupling module, used to perform preference decoupling processing on the semantic features of sample items in M ​​sample domains, and obtain sample domain shared features and sample domain specific features for each sample item, wherein the sample domain shared features represent features shared across domains, and the sample domain specific features represent features specific to the domain, and M≥2; a preference fusion module, used to perform feature fusion processing based on the sample domain shared features and sample domain specific features of each sample item, and obtain sample fusion features for each sample item; a second prediction module, used to predict the recommendation probability of each sample item based on sample user features and the sample fusion features of each sample item, wherein the sample user features represent features of users who interact with the sample items in the M sample domains; and an iterative adjustment module, used to determine the recommendation loss information of the recommendation model based on the recommendation probability of each sample item, and iteratively adjust the network parameters of the recommendation model based on the recommendation loss information until the recommendation loss information meets the recommendation iteration stopping condition, and the network parameters obtained when the recommendation iteration stopping condition is met are used as the trained recommendation model.

[0022] Another aspect of the present disclosure provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above method.

[0023] Another aspect of the present disclosure provides a computer-readable storage medium storing computer-executable instructions, which are used to implement the above method when executed.

[0024] Another aspect of the present disclosure provides a computer program product, which includes computer executable instructions. When the instructions are executed, they are used to implement the above method.

[0025] According to the embodiments of the present disclosure, in response to the operation of the target user in the target application, the target user information of the target user and the item information of at least one item to be recommended in the target application are obtained; then the item information of at least one item to be recommended is subjected to preference decoupling processing to obtain the target domain shared features and target domain specific features of each item to be recommended; then the target domain enhanced features corresponding to the target domain are determined according to the target domain to which the target application belongs; then the recommendation probability of each recommended item is predicted according to the target user information, the target domain shared features, target domain specific features and target domain enhanced features of each item to be recommended; finally, according to the recommendation probability of at least one item to be recommended, a technical means for generating an item recommendation list for the target user is adopted. Since the domain specific features and domain shared features are obtained by performing preference decoupling processing on the item information of the recommended items, and then the target domain enhanced features of the target domain are fused with the domain shared features, only the features that are beneficial to the target domain are fused, thereby reducing the influence of negative information on the recommendation results and improving the accuracy of the recommendation. Therefore, the technical problem that both favorable and unfavorable potential factors are directly migrated during multi-domain recommendation, thereby affecting the accuracy of multi-domain recommendation, is at least partially overcome. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The above and other objects, features and advantages of the present disclosure will become more apparent through the following description of the embodiments of the present disclosure with reference to the accompanying drawings, in which:

[0027] Figure 1 An exemplary system architecture to which the item recommendation method according to an embodiment of the present disclosure can be applied is schematically shown;

[0028] Figure 2 A flowchart of an item recommendation method according to an embodiment of the present disclosure is schematically shown;

[0029] Figure 3 The schematic diagram shows a principle diagram of an item recommendation method according to an embodiment of the present disclosure;

[0030] Figure 4 A flowchart of a training method for a recommendation model according to an embodiment of the present disclosure is schematically shown;

[0031] Figure 5 A schematic diagram schematically shows a semantic extraction based on a semantic extraction module according to an embodiment of the present disclosure;

[0032] Figure 6 A schematic diagram schematically shows preference decoupling based on a preference decoupling module according to an embodiment of the present disclosure;

[0033] Figure 7 A schematic diagram of performing graph convolution processing on domain-specific features according to an embodiment of the present disclosure is schematically shown;

[0034] Figure 8 A schematic diagram schematically illustrates a method for determining sample domain enhancement features according to an embodiment of the present disclosure;

[0035] Fig. 9 The schematic diagram of clustering results according to the embodiment of the present disclosure is schematically shown;

[0036] Fig.10 A block diagram of an item recommendation device according to an embodiment of the present disclosure is schematically shown;

[0037] Fig.11 A block diagram schematically shows a training device for a recommendation model according to an embodiment of the present disclosure; and

[0038] Fig.12 A block diagram of an electronic device suitable for implementing the method described above according to an embodiment of the present disclosure is schematically shown. DETAILED DESCRIPTION

[0039] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0040] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0041] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0042] When using expressions such as "at least one of A, B, and C, etc.", they should generally be interpreted according to the meaning of the expression commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0043] Explanation of terms

[0044] Click-through rate (CTR): refers to the ratio of the number of times a specified content on a website or application (APP) is clicked and exposed. Click-through rate is usually an important indicator for measuring recommendation efficiency in recommendation systems.

[0045] Mutual Information (MI): Mutual information is a measure of the amount of shared information between two variables. It measures how much the uncertainty about one variable is reduced by knowing the value of the other variable. In statistics and information theory, mutual information is used to quantify the strength of the dependency between two random variables. In machine learning, mutual information is often used for feature selection because it can help identify the features that are most correlated with the target variable.

[0046] Loss function: A function used to measure the gap between the prediction results of the recommendation model and the actual results. Usually, the loss function is used to measure the gap between the model's predicted results and the actual ratings, and then optimize the model's parameters so that the model can more accurately predict the user's ratings or purchase behavior for items.

[0047] Feature generalization: Feature generalization is a crucial technique in machine learning and refers to the ability of a model to perform well on unknown data outside of the training data. That is, feature generalization allows the model to make accurate predictions on inputs that were not seen during the training data. It is essential for building robust and scalable models.

[0048] In the embodiments of the present disclosure, the collection, updating, analysis, processing, use, transmission, provision, disclosure, storage, etc. of the data involved (for example, including but not limited to user personal information) are in compliance with the provisions of relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures are taken for user personal information to prevent illegal access to user personal information data and maintain the security of user personal information and network security.

[0049] In the embodiments of the present disclosure, the user's authorization or consent is obtained before obtaining or collecting the user's personal information.

[0050] Recommendation systems are introduced to solve the task of quickly retrieving the most suitable items from a large candidate set to the corresponding user. The complexity of this task comes not only from the massive information in the three stages of the recommendation problem (recall, ranking, and re-ranking), but also from different types of recommendation scenarios.

[0051] Recommendation systems in related technologies generally focus on single-domain recommendations (SDR), but large-scale commercial platforms containing products from multiple domains are becoming more and more common, where the user / item set overlaps to a high degree. Different domains may correspond to different products, and their domains have certain commonalities. Therefore, it is necessary to propose a model to solve cross-domain recommendations (CDR), with the goal of transferring knowledge from a source domain to one or more target domains.

[0052] The main problem facing cross-domain recommendation is that each domain needs to train or fine-tune different models for each source-target domain, but large commercial platforms may consider using a single model to improve the performance of all domains at the same time. This challenge is called multi-domain recommendation (MDR), which aims to extract key information features from multiple domains such as e-commerce, video, and mini-programs, so that they can be applied to the target domain.

[0053] In related technologies, a common method for multi-domain recommendation is to combine multi-domain behaviors as input features. For example, Hao et al. proposed an adversarial feature translation model, which can learn feature translation between different domains under the framework of generative adversarial networks, and can explicitly model the relationship between general and specific domain representations of projects, domains, and users. For another example, Alejandro et al. proposed a structured graph neural network for multi-domain representation learning of sequential user interactions, which can better extract relationships in multiple domains and establish domain-guided representations of recent user interests and representations for long-term interests.

[0054] However, this method combines user behaviors in multiple fields as input features, pays less attention to the interaction of inter-domain features, and fails to disentangle latent factors. This may lead to the direct migration of both favorable and unfavorable latent factors in multi-field recommendations, resulting in low accuracy in multiple fields.

[0055] In response to the above technical problems, an embodiment of the present disclosure provides an item recommendation method, including: in response to an operation of a target user in a target application, obtaining target user information of the target user and item information of at least one item to be recommended in the target application; performing preference decoupling processing on the item information of at least one item to be recommended to obtain target domain shared features and target domain specific features of each item to be recommended; determining target domain enhanced features corresponding to the target domain according to the target domain to which the target application belongs; predicting the recommendation probability of each recommended item according to the target user information, the target domain shared features, the target domain specific features and the target domain enhanced features of each item to be recommended; and generating an item recommendation list for the target user according to the recommendation probability of at least one item to be recommended.

[0056] Figure 1The exemplary system architecture 100 to which the item recommendation method according to the embodiment of the present disclosure can be applied is schematically shown. It should be noted that: Figure 1 What is shown is merely an example of a system architecture to which the embodiments of the present disclosure can be applied, in order to help those skilled in the art understand the technical content of the present disclosure, but it does not mean that the embodiments of the present disclosure cannot be used in other devices, systems, environments or scenarios.

[0057] like Figure 1 As shown, the system architecture 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 is used to provide a medium for a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0058] The user may use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 through the network 104 to receive or send messages, etc. Various communication client applications may be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (only for example).

[0059] The first terminal device 101, the second terminal device 102, and the third terminal device 103 may be various electronic devices having display screens and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0060] The server 105 may be a server that provides various services, such as a background management server (only as an example) that provides support for websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103. The background management server may analyze and process the received data such as user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal device.

[0061] It should be noted that the item recommendation method provided in the embodiment of the present disclosure can generally be executed by the server 105. Accordingly, the item recommendation device provided in the embodiment of the present disclosure can generally be set in the server 105. The item recommendation method provided in the embodiment of the present disclosure can also be executed by a server or server cluster that is different from the server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105. Correspondingly, the item recommendation device provided in the embodiment of the present disclosure can also be set in a server or server cluster that is different from the server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105. Alternatively, the item recommendation method provided in the embodiment of the present disclosure can also be executed by the first terminal device 101, the second terminal device 102 or the third terminal device 103, or can also be executed by other terminal devices different from the first terminal device 101, the second terminal device 102 or the third terminal device 103. Accordingly, the item recommendation device provided in the embodiment of the present disclosure may also be arranged in the first terminal device 101, the second terminal device 102 or the third terminal device 103, or in other terminal devices different from the first terminal device 101, the second terminal device 102 or the third terminal device 103.

[0062] For example, the item information of the item to be recommended may be originally stored in any one of the first terminal device 101, the second terminal device 102 or the third terminal device 103 (for example, the first terminal device 101, but not limited thereto), or stored in an external storage device and imported into the first terminal device 101. Then, the first terminal device 101 may locally execute the item recommendation method provided by the embodiment of the present disclosure, or send the item information of the item to be recommended to other terminal devices, servers, or server clusters, and the other terminal devices, servers, or server clusters that receive the item information of the item to be recommended may execute the item recommendation method provided by the embodiment of the present disclosure.

[0063] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0064] Figure 2 The flowchart of the item recommendation method according to an embodiment of the present disclosure is schematically shown.

[0065] like Figure 2 As shown, the method includes operations S210 to S250.

[0066] In operation S210 , in response to an operation of a target user in a target application, target user information of the target user and item information of at least one item to be recommended in the target application are acquired.

[0067] In operation S220, preference decoupling processing is performed on item information of at least one item to be recommended to obtain target domain shared features and target domain specific features of each item to be recommended.

[0068] In operation S230 , according to the target domain to which the target application belongs, a target domain enhancement feature corresponding to the target domain is determined.

[0069] In operation S240 , the recommendation probability of each recommended item is predicted according to the target user information, the target domain shared features, the target domain specific features, and the target domain enhanced features of each to-be-recommended item.

[0070] In operation S250, an item recommendation list for a target user is generated according to the recommendation probability of each of the at least one item to be recommended.

[0071] According to an embodiment of the present disclosure, the operation of the target user in the target application may include the target user opening the target application, and may also include the target user searching for an item in the target application.

[0072] According to an embodiment of the present disclosure, the at least one item to be recommended may include: all items included in the target application, or part of the items.

[0073] For example, when the target user operates the target application by opening the target application, all items in the target application can be obtained as items to be recommended. For another example, when the target user operates the target application by entering a search keyword to search for items, items related to the search keyword can be obtained as items to be recommended.

[0074] According to the embodiments of the present disclosure, the recommended items refer to objects to be recommended, which can be physical items or virtual items. For example, the recommended items can be physical items such as books, artworks, food, clothes, etc.; the recommended items can also be virtual items such as videos and mini-programs.

[0075] According to an embodiment of the present disclosure, item information may include any information related to the item to be recommended, for example, item information may include item name, item type, item function, etc. Specifically in the field of e-commerce, item information may include information such as product brand and product category.

[0076] According to an embodiment of the present disclosure, the target user information may include basic information of the target user, such as gender, age, etc.

[0077] It should be noted that in the embodiments of the present disclosure, the user's authorization or consent is obtained before obtaining or collecting the user's personal information. For example, before executing operation S210, the authorization information can be displayed to the target user, and after obtaining the target user's authorization or consent, the target user information is obtained.

[0078] According to an embodiment of the present disclosure, target domain shared features represent features shared across domains; target domain specific features represent features specific to a domain.

[0079] According to an embodiment of the present disclosure, performing preference decoupling processing on item information of at least one item to be recommended to obtain the target domain shared features and target domain specific features of each item to be recommended may include: performing preference decoupling processing on the item information of item A to be recommended to obtain the target domain shared features A1 and the target domain specific features A2 of item A to be recommended; performing preference decoupling processing on the item information of item B to be recommended to obtain the target domain shared features B1 and the target domain specific features B2 of item B to be recommended.

[0080] According to an embodiment of the present disclosure, performing preference decoupling processing on item information of at least one item to be recommended to obtain target domain shared features and target domain specific features of each item to be recommended may include: inputting the item information of at least one item to be recommended into a preference decoupling module of a trained recommendation model to perform preference decoupling processing, and outputting the target domain shared features and target domain specific features of each item to be recommended.

[0081] According to an embodiment of the present disclosure, the target domain enhanced feature representation may be appended to the item information during the item recommendation process to enhance the features of the item information.

[0082] According to an embodiment of the present disclosure, determining the target domain enhancement features corresponding to the target domain according to the target domain to which the target application belongs includes: determining the target domain enhancement features for the target domain from a domain-feature relationship graph according to the target domain to which the target application belongs, wherein the domain-feature relationship graph includes at least one domain and the domain enhancement features corresponding to each domain.

[0083] For example, the domain-feature relationship diagram includes domain F1 and the domain enhancement feature corresponding to domain F1, domain F2 and the domain enhancement feature corresponding to domain F2, ..., domain Fn and the domain enhancement feature corresponding to domain Fn.

[0084] According to an embodiment of the present disclosure, the domain-feature relationship graph may be obtained by sharing features among domains of multiple domains.

[0085] In some embodiments, the domain-feature relationship graph can be obtained by clustering the domain shared features of objects in each domain to obtain at least one cluster; then, based on the cluster features of each cluster, a cluster-level cross-domain graph is generated, wherein the vertices of the cluster-level cross-domain graph correspond to the clusters, and the edges represent the similarities between clusters; then, based on the cluster-level cross-domain graph and each cluster, the domain enhancement features of each domain are determined.

[0086] According to an embodiment of the present disclosure, predicting the recommendation probability of each recommended item based on target user information, target domain shared features, target domain specific features and target domain enhanced features of each item to be recommended may include: inputting the target user information, the target domain shared features, target domain specific features and target domain enhanced features of each item to be recommended into a prediction module of a trained recommendation model, and outputting the recommendation probability of each recommended item.

[0087] For example, generating an item recommendation list for a target user according to the recommendation probability of at least one item to be recommended may include: selecting items to be recommended whose recommendation probability is greater than a preset threshold to generate the item recommendation list.

[0088] For example, generating an item recommendation list for a target user based on the recommendation probability of at least one item to be recommended may include: sorting at least one item to be recommended based on the recommendation probability, and selecting an item to be recommended whose sorting position is before a preset position to generate an item recommendation list.

[0089] According to the embodiments of the present disclosure, a technical means is adopted to obtain the target user information of the target user and the item information of at least one item to be recommended in the target application in response to the operation of the target user in the target application; then the item information of at least one item to be recommended is subjected to preference decoupling processing to obtain the target domain shared features and target domain specific features of each item to be recommended; then the target domain enhanced features corresponding to the target domain are determined according to the target domain to which the target application belongs; then the recommendation probability of each recommended item is predicted according to the target user information, the target domain shared features, target domain specific features and target domain enhanced features of each item to be recommended; finally, according to the recommendation probability of at least one item to be recommended, a list of recommended items for the target user is generated. Since the domain specific features and domain shared features are obtained by performing preference decoupling processing on the item information of the items to be recommended, and then the target domain enhanced features of the target domain are fused with the domain shared features, only the features that are beneficial to the target domain are fused, thereby reducing the impact of negative information on the recommendation results and improving the accuracy of recommendations. Therefore, the technical problem of directly migrating both favorable and unfavorable potential factors during multi-domain recommendations, thereby affecting the accuracy of multi-domain recommendations, is at least partially overcome.

[0090] According to an embodiment of the present disclosure, the above-mentioned item recommendation method also includes: performing semantic feature extraction on the target user information and the item information of at least one item to be recommended to obtain the target user feature and the item feature to be recommended of each item to be recommended; wherein, performing preference decoupling processing on the item information of at least one item to be recommended to obtain the target domain shared feature and the target domain specific feature of each item to be recommended includes: performing preference decoupling processing on the item feature to be recommended of each item to be recommended to obtain the target domain shared feature and the target domain specific feature of each item to be recommended; predicting the recommendation probability of each recommended item according to the target user information, the target domain shared feature, the target domain specific feature and the target domain enhanced feature of each item to be recommended includes: predicting the recommendation probability of each recommended item according to the target user feature, the target domain shared feature, the target domain specific feature and the target domain enhanced feature of each item to be recommended.

[0091] According to an embodiment of the present disclosure, performing semantic feature extraction on target user information and item information of at least one item to be recommended may include extracting key information from the target user information and converting the key information into a vector form, and at the same time, extracting key information from the item information of the item to be recommended and converting it into a vector form, so that the preference decoupling module can capture domain-specific features and domain-shared features.

[0092] According to an embodiment of the present disclosure, semantic feature extraction is performed on target user information and item information of at least one item to be recommended to obtain target user features and item features of each item to be recommended, which may include: inputting the target user information and item information of at least one item to be recommended into a semantic extraction module of a trained recommendation model to perform semantic feature extraction, and outputting the target user features and item features of each item to be recommended.

[0093] According to an embodiment of the present disclosure, preference decoupling processing is performed on the item information of at least one item to be recommended to obtain the target domain shared features and target domain specific features of each item to be recommended, including: performing domain sharing encoding processing on the item information of at least one item to be recommended to obtain the target domain shared features of each recommended item; performing domain specific encoding processing on the item information of at least one item to be recommended to obtain the target domain specific features of each recommended item.

[0094] According to an embodiment of the present disclosure, performing domain sharing encoding processing on item information of at least one item to be recommended to obtain target domain shared features of each recommended item may include: inputting the item information of at least one item to be recommended into a trained domain sharing encoder, and outputting the target domain shared features of each recommended item.

[0095] According to an embodiment of the present disclosure, performing domain-specific encoding processing on item information of at least one item to be recommended to obtain target domain-specific features of each recommended item may include: inputting the item information of at least one item to be recommended into respective corresponding domain-specific encoders, and outputting target domain-specific features of each recommended item.

[0096] According to an embodiment of the present disclosure, predicting the recommendation probability of each recommended item based on target user information, target domain shared features, target domain specific features and target domain enhanced features of each item to be recommended includes: for each item to be recommended, determining the target fusion features of each item to be recommended based on the target domain shared features, target domain specific features and target domain enhanced features; predicting the recommendation probability of each recommended item based on the target user information and the target fusion features of each item to be recommended.

[0097] According to an embodiment of the present disclosure, determining the target fusion features of each to-be-recommended item based on the target domain shared features, the target domain specific features, and the target domain enhanced features may include: combining the target domain shared features, the target domain specific features, and the target domain enhanced features to obtain the target fusion features.

[0098] Figure 3 The schematic diagram shows a principle diagram of an item recommendation method according to an embodiment of the present disclosure.

[0099] like Figure 3 As shown, the item recommendation method in this embodiment may include: when it is detected that the target user operates in the target application, such as opening the target application, obtaining item information 310, target user information 330 and target application information 320 of at least one item to be recommended in the target application, wherein the item information 310 of the item to be recommended may include item information 1 to be recommended, item information 2 to be recommended, ..., item information n to be recommended.

[0100] For example, the item information 310 and the target user information 330 of the item to be recommended can be respectively input into the semantic extraction module of the trained recommendation model for semantic feature extraction, and the feature 311 of the item to be recommended and the target user feature 331 can be output, wherein the feature 311 of the item to be recommended includes feature 1 of the item to be recommended, feature 2 of the item to be recommended, ..., feature n of the item to be recommended, and feature 1 of the item to be recommended corresponds to information 1 of the item to be recommended, feature 2 of the item to be recommended corresponds to information 2 of the item to be recommended, ..., feature n of the item to be recommended corresponds to information n of the item to be recommended.

[0101] For example, the feature 311 of the item to be recommended can be input into the preference decoupling module of the trained recommendation model for preference decoupling processing, and the decoupling result 312 is output, wherein the decoupling result 312 includes target domain shared feature 1 and target domain feature feature 1, target domain shared feature 2 and target domain feature feature 2, ..., target domain shared feature n and target domain feature feature n, wherein target domain shared feature 1 and target domain feature feature 1 correspond to feature 1 of the item to be recommended, target domain shared feature 2 and target domain feature feature 2 correspond to feature 2 of the item to be recommended, ..., target domain shared feature n and target domain feature feature n correspond to feature n of the item to be recommended.

[0102] For example, the decoupling result 312 can be input into the preference fusion module of the trained recommendation model. The preference fusion module determines the target domain enhanced feature 321 according to the target application information 320 and the preset domain and feature relationship diagram, and performs feature fusion on the target domain enhanced feature 321 and the decoupling result 312 to obtain a fusion result 350. The fusion result 350 includes target fusion feature 1, target fusion feature 2, ..., target fusion feature n, and target fusion feature 1 is obtained by fusion of target domain shared feature 1, target domain feature feature 1 and target domain enhanced feature; target fusion feature 2 is obtained by fusion of target domain shared feature 2, target domain feature feature 2 and target domain enhanced feature; ...; target fusion feature n is obtained by fusion of target domain shared feature n, target domain feature feature n and target domain enhanced feature.

[0103] For example, the decoupling result 350 and the target user feature 331 can be input into the prediction module of the trained recommendation model to predict the recommendation probability, and output the prediction result 360, which includes the recommendation probability of the to-be-recommended item 1, the recommendation probability of the to-be-recommended item 2, ..., the recommendation probability of the to-be-recommended item n. Then, a recommendation list 370 is generated according to the recommendation probability of each to-be-recommended item.

[0104] Figure 4 The flowchart of the training method of the recommendation model according to the embodiment of the present disclosure is schematically shown.

[0105] like Figure 4 As shown, the training method of the recommendation model of this embodiment may include operations S410 to S440.

[0106] In operation S410, preference decoupling is performed on the semantic features of sample items in the M sample domains to obtain sample domain shared features and sample domain specific features for each sample item. The sample domain shared features represent features shared across domains, and the sample domain specific features represent features specific to a domain, M≥2.

[0107] In operation S420, feature fusion processing is performed based on the sample domain shared features and sample domain specific features of each sample item to obtain a sample fusion feature of each sample item.

[0108] In operation S430, the recommendation probability of each sample item is predicted based on the sample user feature and the sample fusion feature of each sample item, wherein the sample user feature represents the feature of the user who interacts with the sample items in the M sample fields.

[0109] In operation S440, the recommendation loss information of the recommendation model is determined based on the recommendation probability of each sample item, and the network parameters of the recommendation model are iteratively adjusted according to the recommendation loss information until the recommendation loss information meets the recommendation iteration stop condition, and the network parameters obtained when the recommendation iteration stop condition is met are used as the trained recommendation model.

[0110] According to an embodiment of the present disclosure, preference decoupling processing is performed on the semantic feature a1 of the sample item in the sample domain F1 to obtain the sample domain shared feature a11 and the sample domain specific feature a12; preference decoupling processing is performed on the semantic feature b1 of the sample item in the sample domain F2 to obtain the sample domain shared feature b11 and the sample domain specific feature b12.

[0111] According to an embodiment of the present disclosure, the semantic features of the sample items may be obtained by extracting features from the text data of the sample items, for example, inputting the text data of the sample items into a semantic extraction module to extract semantic features and outputting the semantic features of the sample items.

[0112] According to an embodiment of the present disclosure, the M sample fields may include, for example, an e-commerce field, a video field, a mini program field, etc.

[0113] According to an embodiment of the present disclosure, the sample user-sample item interactions may include items purchased by the user, items clicked by the user, items viewed by the user, and the like.

[0114] According to an embodiment of the present disclosure, determining the recommendation loss information of the recommendation model based on the recommendation probability of each sample item may include determining the loss value using a preset loss function according to the recommendation probability of each sample item and the historical click rate of the sample item.

[0115] According to an embodiment of the present disclosure, the recommendation model also includes a semantic extraction module; the method also includes: obtaining text data of sample items in M ​​sample fields and sample user information of users who interact with the sample items in the M sample fields; inputting the text data of the sample items and the sample user information into the semantic extraction module, and outputting sample user features and semantic features of the sample items in each of the M sample fields.

[0116] According to an embodiment of the present disclosure, the text data of the sample item may include any text of the user describing the sample item. For example, the text data may include the name, type, function, etc. of the sample item. The sample user information may include basic information of the sample user, such as gender, age, etc.

[0117] In some of the embodiments, in order to obtain the semantic features of the sample items, the semantic extraction module can use a large language model BERT for semantic extraction to extract word embedding for each word of the text data. In order to protect user privacy and increase the applicability of the method, an ID embedding method can be used. For each sample domain, the user embedding, that is, the sample user feature, is defined as a trainable vector 𝑢, which can be retrieved by the ID of 𝑢. Then, the user embedding in each sample domain can be represented as an embedding matrix, in which each row corresponds to a user embedding.

[0118] It should be noted that in the embodiments of the present disclosure, the user's authorization or consent is obtained before obtaining or collecting the user's personal information. For example, authorization information can be displayed to the sample user before obtaining the sample user information, and the sample user information is obtained after obtaining the sample user's authorization or consent.

[0119] Figure 5 The schematic diagram schematically shows a semantic extraction based on a semantic extraction module according to an embodiment of the present disclosure.

[0120] like Figure 5 As shown, this embodiment includes text data 510 of sample items, for example, text data of sample item 1, text data of sample item 2, ..., text data of sample item n; the text data 510 of the sample items is input into a semantic extraction module 520 for semantic extraction, and sample item semantic features 530 are output respectively, which may include semantic features of sample item 1, semantic features of sample item 2, ..., semantic features of sample item n.

[0121] According to an embodiment of the present disclosure, the recommendation model includes a preference decoupling module, which includes a domain-shared encoder and a domain-specific encoder corresponding to each of the M sample domains; wherein, preference decoupling processing is performed on the semantic features of the sample items of each of the M sample domains to obtain the sample domain-shared features and sample domain-specific features of each sample item, including: inputting the semantic features of the sample items of each of the M sample domains into the domain-shared encoder, and outputting the sample domain-shared features of each sample item; inputting the semantic features of the sample items of each of the M sample domains into the corresponding domain-specific encoders, and outputting the sample domain-specific features of each sample item.

[0122] Figure 6The diagram schematically shows preference decoupling based on a preference decoupling module according to an embodiment of the present disclosure.

[0123] like Figure 6 As shown, this embodiment includes semantic features of sample items in sample domain FA610 and sample domain FB620, wherein sample domain FA610 includes semantic features of sample item A1, semantic features of sample item A2, ..., semantic features of sample item An; sample items in sample domain FB610 include semantic features of sample item B1, semantic features of sample item B2, ..., semantic features of sample item Bn. Then, the semantic features of sample items in sample domain FA610 and sample domain FB620 are input into preference decoupling module 630 for preference decoupling processing, for example, the semantic features of sample items in sample domain FA610 are input into domain-shared encoder and domain-specific encoder A in preference decoupling module 630, and sample domain-shared features 640 and sample domain-specific features 660 for sample domain FA610 are output respectively; then, the semantic features of sample items in sample domain FB620 are input into domain-shared encoder and domain-specific encoder B in preference decoupling module 630, and sample domain-shared features 650 and sample domain-specific features 670 for sample domain FB620 are output respectively. Among them, the sample domain shared features 640 include the sample domain shared features of sample item A1, the sample domain shared features of sample item A2, ..., the sample domain shared features of sample item An; the sample domain specific features 660 include the sample domain specific features of sample item A1, the sample domain specific features of sample item A2, ..., the sample domain specific features of sample item An; the sample domain shared features 650 include the sample domain shared features of sample item B1, the sample domain shared features of sample item B2, ..., the sample domain shared features of sample item Bn; the sample domain specific features 670 include the sample domain specific features of sample item B1, the sample domain specific features of sample item B2, ..., the sample domain specific features of sample item Bn.

[0124] According to an embodiment of the present disclosure, the training method of the above-mentioned recommendation model also includes: determining the specific feature similarity between every two sample domains using a first preset formula based on the sample domain specific features of each sample item; and adjusting the network parameters of the domain-specific encoder using the specific feature similarity.

[0125] According to an embodiment of the present disclosure, using the first preset formula to determine the specific feature similarity between every two sample fields may include: for two of the sample fields, using the first preset formula to determine the specific feature similarity between every two sample items in the two sample fields, and determining the specific feature similarity between the two sample fields based on the specific feature similarity between every two sample items.

[0126] According to an embodiment of the present disclosure, based on the sample domain specific features of each sample item, determining the specific feature similarity between every two sample domains using the first preset formula may include: determining the specific feature similarity between every two sample domains using formula (1).

[0127] (1).

[0128] Among them, m and n represent sample domain m and sample domain n, represents the first The sample domain-specific features of the sample items, and represents the number of sample items in sample area m and sample area n, It represents the conditional probability, which indicates the possibility of predicting the sample domain specific features of another sample item j given the sample domain specific features of the i-th sample item in the sample domain m, reflecting the similarity relationship between the items in the sample domain m and the sample domain n. When the sample items i and j come from two different sample domains, the domain feature encoders corresponding to the two sample items are adjusted to minimize the conditional probability in the formula. Ability to reduce the mutual dependency of two sample items to optimize domain-specific encoders.

[0129] The first term in formula (1) represents the calculation of mutual information between sample items within the sample domain m, with the purpose of capturing domain-specific information in the sample domain. The second term represents the mutual information calculation between cross-domains m and n, which aims to capture the information or preferences shared across domains.

[0130] in The calculation method can be expressed as the following formula (2).

[0131] (2);

[0132] in, and The encoder is shared by both domains and get.

[0133] (3)

[0134] in, and represent the sample domain-specific features of sample items i and j respectively.

[0135] The contrast log-ratio upper bound (CLUB) is a differentiable approximate upper bound of the mutual information, and thus can be minimized using a gradient descent algorithm. However, CLUB is only applicable to domain-shared users and items, because it estimates the mutual information between two representations of the same object, such as two embedding types of users or items, which is not available in non-overlapping multi-domain recommendations. Therefore, the disclosed embodiment proposes AFCLUB, a novel anchor-free CLUB variant that can measure the mutual information between all items in different domains to remove the identification of domain-shared items, which can be expressed as the above formula (1).

[0136] According to an embodiment of the present disclosure, based on the sample domain shared features of each sample item, a second preset formula is used to determine the shared feature similarity between every two sample domains; and the network parameters of the domain shared encoder are adjusted using the shared feature similarity.

[0137] According to an embodiment of the present disclosure, using a second preset formula to determine the shared feature similarity between every two sample fields may include: for two of the sample fields, using the second preset formula to determine the shared feature similarity of every two sample items in the two sample fields, and determining the shared feature similarity between the two sample fields based on the shared feature similarity of every two sample items.

[0138] According to an embodiment of the present disclosure, based on the sample domain shared features of each sample item, the similarity of shared features between every two sample domains is determined using formula (4).

[0139] (4).

[0140] in, represents the anchor-independent domain contrast loss function between sample domain m and sample domain n, which is used to maximize the shared information between different domains. represents the cosine similarity, , represents the inner product of x and y, ||x|| represents the L2 norm of x, represents the exponential function with base e, represents the logarithmic function with base e, m and n represent sample domain m and sample domain n, represents the first The sample domain of sample items shares features, and Indicates the number of sample items in sample domain m and sample domain n. Hyperparameter is the normalization parameter,

[0141] By using formula (4) to determine the shared feature similarity between every two sample domains, the network parameters of the domain-sharing encoder are adjusted using the shared feature similarity to minimize the above loss function, which can maximize the lower bound of the anchor-free domain contrast estimation (AFDCE) and maximize the mutual information between domain-sharing embeddings, thereby extracting better domain-sharing preferences.

[0142] Information-Noise Contrastive Estimation (InfoNCE) is a widely used lower bound for estimating the mutual information between two variables. However, the measurement of InfoNCE requires the use of domain-shared users or items as anchors in multi-domain recommendations. Therefore, the present disclosure proposes an anchor-free domain contrastive estimation (AFDCE), which can measure the mutual information between domain-shared items embedded in each domain without using domain-shared users. The inverse AFDCE loss can be expressed by formula (4).

[0143] According to an embodiment of the present disclosure, the training method of the recommendation model may further include: for each sample domain, generating a specific feature interaction graph for the sample domain according to the sample domain specific features and sample user features of each sample item, then inputting the specific feature interaction graph into the first graph convolution layer, aggregating the interaction information into the specific feature interaction graph, and outputting the sample domain specific features after graph convolution, wherein in this specific feature interaction graph, each vertex corresponds to all sample users and sample items in the sample domain, and the edge is the interaction between the sample user and the sample item in the sample domain. Then, according to the sample domain specific features after graph convolution of each sample item, the specific feature similarity between each two sample domains is determined using formula (1); and the network parameters of the domain specific encoder are adjusted using the specific feature similarity.

[0144] Figure 7 A schematic diagram of performing graph convolution processing on domain-specific features according to an embodiment of the present disclosure is schematically shown.

[0145] like Figure 7 As shown, in this embodiment, for the sample domain FA, its sample item domain-specific features 710 include sample domain-specific features of sample item A1, sample domain-specific features of sample item A2, ..., sample domain-specific features of sample item An, and a specific feature interaction graph 730 is generated according to the sample item domain-specific features 710 and the sample user features A720; then, graph convolution processing is performed on the specific feature interaction graph 730 to obtain domain-specific features 740 after graph convolution, and the domain-specific features 740 after graph convolution include graph convolution specific features of sample item A1, graph convolution specific features of sample item A2, ..., graph convolution specific features of sample item An.

[0146] According to an embodiment of the present disclosure, the training method of the recommendation model may further include: for each sample domain, generating a shared feature interaction graph for the sample domain according to the sample domain shared features and sample user features of each sample item, then inputting the shared feature interaction graph into the second graph convolution layer, aggregating the interaction information into the shared feature interaction graph, and outputting the sample domain shared features after graph convolution, wherein in this shared feature interaction graph, each vertex corresponds to all sample users and sample items in the sample domain, and the edge is the interaction between the sample user and the sample item in the sample domain. Then, according to the sample domain shared features after graph convolution of each sample item, the shared feature similarity between each two sample domains is determined using formula (4); and the network parameters of the domain shared encoder are adjusted using the shared feature similarity.

[0147] The process of performing graph convolution on domain-shared features is similar to Figure 7 The method is similar and will not be described again here.

[0148] According to an embodiment of the present disclosure, a sample domain shared encoder and a set of sample domain specific encoders are used to encode sample semantic features to generate sample domain specific and sample domain shared item features. Then, the mutual information of these two types of features is measured to optimize these sample domain shared encoders and sample domain specific encoders. The optimized sample domain shared encoder and sample domain specific encoder will generate features that contain sample domain shared and sample domain specific. Using this approach can alleviate the negative transfer problem and improve recommendations. Afterwards, a graph convolution layer is used to aggregate the mutual information into sample domain specific and sample domain shared features. The mutual information can be obtained from the interaction graph of each domain. Since no additional parameters are introduced, the graph convolution layer can effectively alleviate the overfitting problem of sparse domains. Finally, based on the two types of item embeddings generated by the two types of encoders, namely sample domain shared features and sample domain specific features, constraints based on mutual information are used to decouple domain specific and domain shared preferences.

[0149] According to an embodiment of the present disclosure, the recommendation model includes a preference fusion module; feature fusion processing is performed according to the sample domain shared features and sample domain specific features of each sample item to obtain the sample fusion features of each sample item, including: inputting the sample domain shared features and sample domain specific features of each sample item into the preference fusion module, and determining the sample domain enhanced features of each sample item according to the sample domain shared features of each sample item; feature fusion processing is performed according to the sample domain shared features, sample domain specific features and sample domain enhanced features of each sample item, and the sample fusion features of each sample item are output.

[0150] According to an embodiment of the present disclosure, the sample domain enhanced feature representation can be transferred to the features of the sample domain.

[0151] In some embodiments, determining the sample domain enhanced features of each sample object according to the sample domain shared features of each sample object may include: fusing the sample domain shared features of each sample object to obtain the sample domain enhanced features.

[0152] For example, the sample domain shared features of each sample item are directly fused using a fusion encoder to obtain sample domain enhanced features. In this case, the sample domain enhanced features corresponding to each sample domain are the same.

[0153] In other embodiments, determining the sample domain enhanced features of each sample field based on the sample domain shared features of each sample item may include: clustering the sample domain shared features of each sample item to obtain at least one cluster cluster; generating a cluster-level cross-domain graph based on the cluster features of each cluster cluster, wherein the vertices of the cluster-level cross-domain graph correspond to the cluster clusters and the edges represent the similarities between clusters; and determining the sample domain enhanced features of each sample field based on the cluster-level cross-domain graph and each cluster cluster.

[0154] According to an embodiment of the present disclosure, each cluster includes at least one sample domain shared feature. Generating a cluster-level cross-domain graph according to the cluster features of each cluster includes: for each cluster, determining the cluster feature according to the sample domain shared feature contained in the cluster.

[0155] For example, all sample domain shared features in the cluster are multiplied and averaged to obtain the cluster feature.

[0156] According to an embodiment of the present disclosure, the edges in the cluster-level cross-domain graph are determined by the similarity between vertex clusters. For example, the similarity between cluster i and cluster j is It can be obtained by the following formula (5)

[0157] (5);

[0158] Where i and j represent cluster i and cluster j, and Represents the cluster features of cluster i and cluster j. The softmax function converts an arbitrary d-dimensional vector σ(z) x into another real number vector ,in The j-th dimension of can be expressed as , represents the exponential function with base e, represents the cosine similarity, , represents the inner product of x and y, and ||x|| represents the L2 norm of x.

[0159] Figure 8A schematic diagram schematically illustrates a method for determining sample domain enhancement features according to an embodiment of the present disclosure.

[0160] like Figure 8 As shown, the embodiment includes a sample domain shared feature 810 of a sample domain FA and a sample domain shared feature 820 of a sample domain B, and then adaptively clusters the sample domain shared features of sample item A1, sample domain shared features of sample item A2, ..., sample domain shared features of sample item An in the sample domain shared features 810 and the sample domain shared features of sample item B1, sample domain shared features of sample item B2, ..., sample domain shared features of sample item Bn' in the sample domain shared features 820 to obtain cluster clusters a830 and cluster clusters b840. b840, then, the cluster features of cluster a are determined according to the sample domain shared features a1, sample domain shared features a2, ..., sample domain shared features am; the cluster features of cluster b are determined according to the sample domain shared features b1, sample domain shared features b2, ..., sample domain shared features bn'; then, a cluster-level cross-domain graph 850 is constructed according to the cluster features of cluster a and the cluster features of cluster b; based on the cluster-level cross-domain graph 850 and cluster a and cluster b, the sample domain enhanced features 860 of sample domain FA and the sample domain enhanced features 870 of sample domain FB are determined.

[0161] Fig. 9 A schematic diagram of clustering results according to an embodiment of the present disclosure is schematically shown.

[0162] like Fig. 9 As shown, after clustering the sample domain shared features 910 of the sample items, a clustering result 920 is obtained, and the clustering result 920 includes four clusters, namely cluster 1, cluster 2, cluster 3 and cluster 4.

[0163] According to an embodiment of the present disclosure, determining the sample domain enhanced features of each sample domain according to the cluster-level cross-domain graph and each clustering cluster includes: performing domain sharing preference aggregation processing on the cluster-level cross-domain graph to obtain enhanced cluster features of each clustering cluster, wherein the enhanced cluster feature characterization adds inter-domain shared features to the cluster features; determining the sample domain enhanced features of each sample domain according to the cluster features and enhanced cluster features of each clustering cluster.

[0164] According to an embodiment of the present disclosure, performing domain sharing preference aggregation processing on the cluster-level cross-domain graph to obtain enhanced cluster features for each cluster cluster may include: inputting the cluster-level cross-domain graph into the third graph convolution layer, fusing and transferring cross-domain relationship knowledge, and aggregating domain sharing features of different sample fields into cluster features to obtain enhanced cluster features.

[0165] According to an embodiment of the present disclosure, determining the sample domain enhanced features of each sample field according to the cluster features and enhanced cluster features of each cluster cluster includes: for the jth sample field among the M sample fields, determining the jth cluster allocation matrix of the jth sample field according to the cluster features of each cluster cluster, wherein the jth cluster allocation matrix includes the allocation coefficient of each cluster cluster in the jth sample field, 1≤j≤M; determining the jth sample domain enhanced features of the jth sample field according to the j cluster allocation matrices and the enhanced cluster features of each cluster cluster.

[0166] According to an embodiment of the present disclosure, the allocation coefficient of each clustering cluster in the j-th sample domain may include: for one of the clustering clusters, determining the sample domain shared features belonging to the j-th sample domain in the clustering cluster, and determining the allocation coefficient according to the number of sample domain shared features belonging to the j-th sample domain and the number of sample domain shared features in the clustering cluster.

[0167] For example, the cluster includes sample domain shared features 1, 2, 3, 4, and 5, wherein the sample domain shared features belong to the jth sample domain, and the allocation coefficient of the cluster in the jth sample domain may be 1 / 5=0.2.

[0168] According to an embodiment of the present disclosure, determining the j-th sample domain enhanced feature of the j-th sample domain based on j cluster allocation matrices and the enhanced cluster features of each cluster cluster includes: for each cluster cluster, multiplying the enhanced cluster feature of the cluster cluster with the allocation coefficient of the cluster cluster in the j-th sample domain to obtain the allocation feature of the cluster cluster; and summing the allocation features of each cluster cluster to obtain the j-th sample domain enhanced feature.

[0169] For example, in some embodiments, cluster 1, cluster 2, cluster 3, and cluster 4 are included; the distribution coefficients of cluster 1, cluster 2, cluster 3, and cluster 4 in the jth sample domain are cc1, cc2, cc3, and cc4 respectively. Then, multiplying the enhanced cluster features of the clusters with the distribution coefficients of the clusters in the jth sample domain to obtain the distribution features of the clusters may include: multiplying the enhanced features of cluster 1 with the distribution system c1 to obtain the distribution feature b1; multiplying the enhanced features of cluster 2 with the distribution system c2 to obtain the distribution feature b2; multiplying the enhanced features of cluster 3 with the distribution system c3 to obtain the distribution feature b3; and multiplying the enhanced features of cluster 4 with the distribution system c4 to obtain the distribution feature b4. Then, b1, b2, b3, and b4 are summed to obtain the jth sample domain enhanced features.

[0170] It should be noted that the j-th sample domain enhancement feature is the sample item enhancement feature of the sample item belonging to the j-th sample domain.

[0171] According to an embodiment of the present disclosure, a domain-feature relationship graph may be obtained based on the sample domain enhancement features of each sample domain.

[0172] According to an embodiment of the present disclosure, the training method of the recommendation model further includes: determining a sample item enhancement feature of each sample item according to a sample domain enhancement feature of each sample domain;

[0173] Determine the enhanced domain shared features of the sample items according to the sample domain shared features and the sample item enhanced features of each sample item;

[0174] For an i-th sample domain among the M sample domains, determining a first preference score for the sample domain specific feature according to sample user features and sample domain specific features of sample items in the i-th sample domain;

[0175] Determine a second preference score for the enhanced domain shared feature according to the sample user feature and the enhanced domain shared feature of the sample item;

[0176] Based on the first preference score and the second preference score, fusion loss information for the preference fusion module is determined, and network parameters of the preference fusion module are iteratively adjusted according to the fusion loss information until the fusion loss information satisfies a fusion iteration stop condition.

[0177] According to an embodiment of the present disclosure, determining fusion loss information for a preference fusion module based on the first preference score and the second preference score may include determining fusion loss information according to a KL divergence between the first preference score and the second preference score based on a preference fusion loss function.

[0178] Among them, the preference fusion loss function is as follows (6).

[0179] (6);

[0180] in, represents the preference fusion loss function on the sample domain d, which is used to measure the difference between domain-specific features and enhanced domain shared features in order to better fuse this information. Represents a sample user and sample items The set of interactions in the sample domain d belongs to This interaction set contains the interaction data between sample users and sample items. Represents a sample user For sample items The true preference label in the sample domain d, Represents the sample user predicted by the recommendation model For sample items The preference scores in sample domain d, such as the first preference score and the second preference score, represents the logarithmic function with base e. The first term in formula (6) Indicates that when the true label When , the cross entropy loss between the predicted value and the true value is calculated to measure the prediction error of the model for the positive example. The second term in formula (6) Indicates that when the real , calculate the prediction error of the model for negative examples.

[0181] According to an embodiment of the present disclosure, sample items are adaptively clustered according to their sample domain shared features, and a cluster-level cross-domain graph is constructed; then, the structural information of the cluster-level cross-domain graph is aggregated through graph convolution to enhance the domain shared features in each sample domain. In this way, domain shared features in multiple sample domains can be integrated into each sample domain without domain-shared users and items. Afterwards, these enhanced domain shared features are combined with domain-specific features to further improve recommendation performance.

[0182] According to an embodiment of the present disclosure, the Bayesian personalized ranking (BPR) loss is used to measure the prediction performance. For multi-domain recommendation problems, the prediction goal is to minimize the BPR loss of all domains simultaneously. By randomly sampling negative sample items for each interaction, the BPR loss of the domain is measured by the following formula (7):

[0183] (7)

[0184] in, Represents the Bayesian personalized ranking loss on the sample domain d, which is used to optimize the recommendation model so that for each sample user, the recommendation model can better rank the sample items that the sample user prefers and dislikes. represents the set of ternary interactions in the sample domain d, where is a sample user, are sample items that sample users have interacted with (positive examples), are sample items that the user has not interacted with (negative examples), Represents the sample user predicted by the recommendation model For sample items Preference score in sample domain d, Represents the sample user predicted by the recommendation model For sample items Preference score in sample domain d, represents the sigmoid function, , represents the logarithmic function with base e. Finally, the loss function L of the entire recommendation model can be expressed as follows (8):

[0185] (8)

[0186] in, , and It is a hyperparameter that controls each component of the loss function. By minimizing the final loss function, the training of the mutual information multi-domain recommendation model can be achieved.

[0187] According to an embodiment of the present disclosure, a recommendation model is proposed for multi-domain recommendation tasks where users have no direct overlap between different domains, and an anchor-free contrast log ratio upper bound (AFCLUB) method is proposed to measure the mutual information between item representations in different domains. By minimizing AFCLUB, general preferences that are not specific to any domain can be learned, such as plots or themes that users may generally like. In addition, an anchor-free domain contrast estimation (AFDCE) method is proposed in the embodiment of the present disclosure as a lower bound on the mutual information between item representations in different domains, and preferences specific to a certain domain are learned by maximizing AFDCE, such as a movie genre or book style that a certain user particularly likes. The embodiment of the present disclosure further provides a cluster-level cross-domain graph to migrate and fuse the learned domain shared preferences between multiple domains, combining the domain shared preferences with the domain-specific preferences in each domain to simultaneously improve the recommendation effect in all domains.

[0188] According to the embodiments of the present disclosure, the proposed technical solution can reduce the impact of negative information in multi-field recommendations, better integrate different preferences, and achieve more accurate multi-field recommendations, with the offline AUC (area under curve) improved by 0.15%.

[0189] Based on the above-mentioned item recommendation method, the present disclosure also provides an item recommendation device. Fig.10 A detailed description of the item recommendation device

[0190] Fig.10 The block diagram of the item recommendation device according to an embodiment of the present disclosure is schematically shown.

[0191] like Fig.10 As shown, the item recommendation device 1100 of this embodiment includes a first acquisition module 1110 , a decoupling processing module 1120 , a first determination module 1130 , a first prediction module 1140 and a generation module 1150 .

[0192] The first acquisition module 1110 is used to obtain target user information of the target user and item information of at least one item to be recommended in the target application in response to the target user's operation in the target application. In one embodiment, the first acquisition module 1110 can be used to perform the operation S210 described above, which will not be repeated here.

[0193] The decoupling processing module 1120 is used to perform preference decoupling processing on the item information of at least one item to be recommended, and obtain the target domain shared features and target domain specific features of each item to be recommended. In one embodiment, the decoupling processing module 1120 can be used to perform the operation S220 described above, which will not be repeated here.

[0194] The first determination module 1130 is used to determine the target domain enhancement feature corresponding to the target domain according to the target domain to which the target application belongs. In one embodiment, the first determination module 1130 can be used to perform the operation S230 described above, which will not be described in detail here.

[0195] The first prediction module 1140 is used to predict the recommendation probability of each recommended item according to the target user information, the target domain shared features, the target domain specific features and the target domain enhanced features of each item to be recommended. In one embodiment, the first prediction module 1140 can be used to perform the operation S240 described above, which will not be repeated here.

[0196] The generating module 1150 is used to generate an item recommendation list for the target user according to the recommendation probability of at least one item to be recommended. In one embodiment, the generating module 1150 can be used to perform the operation S250 described above, which will not be described in detail here.

[0197] According to an embodiment of the present disclosure, the decoupling processing module includes: a first domain shared encoding submodule and a first domain specific encoding submodule.

[0198] The first domain sharing encoding submodule is used to perform domain sharing encoding processing on the item information of at least one item to be recommended to obtain the target domain sharing features of each recommended item.

[0199] The first domain-specific encoding submodule is used to perform domain-specific encoding processing on item information of at least one item to be recommended to obtain target domain-specific features of each recommended item.

[0200] According to an embodiment of the present disclosure, the first determination module includes: a first determination submodule.

[0201] The first determination submodule is used to determine the target domain enhancement features for the target domain from the domain-feature relationship diagram according to the target domain to which the target application belongs, wherein the domain-feature relationship diagram includes at least one domain and the domain enhancement features corresponding to each domain.

[0202] According to an embodiment of the present disclosure, the first prediction module includes: a second determination submodule and a prediction submodule.

[0203] The second determination submodule is used to determine, for each item to be recommended, the target fusion feature of each item to be recommended according to the target domain shared feature, the target domain specific feature and the target domain enhanced feature.

[0204] The prediction submodule is used to predict the recommendation probability of each recommended item based on the target user information and the target fusion features of each item to be recommended.

[0205] According to an embodiment of the present disclosure, the above-mentioned item recommendation device also includes: a feature extraction module.

[0206] The feature extraction module is used to extract semantic features from target user information and item information of at least one item to be recommended, so as to obtain target user features and item features of each item to be recommended.

[0207] According to an embodiment of the present disclosure, the decoupling processing module is also used to perform preference decoupling processing on the features of each item to be recommended, so as to obtain the target domain shared features and target domain specific features of each item to be recommended.

[0208] According to an embodiment of the present disclosure, the first prediction module is also used to predict the recommendation probability of each recommended item based on the target user characteristics, the target domain shared characteristics, the target domain specific characteristics and the target domain enhanced characteristics of each item to be recommended.

[0209] According to an embodiment of the present disclosure, any multiple of the first acquisition module 1110, the decoupling processing module 1120, the first determination module 1130, the first prediction module 1140, and the generation module 1150 can be combined in one module / unit / sub-unit for implementation, or any one of the modules / units / sub-units can be split into multiple modules / units / sub-units. Alternatively, at least part of the functions of one or more of these modules / units / sub-units can be combined with at least part of the functions of other modules / units / sub-units and implemented in one module / unit / sub-unit. According to an embodiment of the present disclosure, at least one of the first acquisition module 1110, the decoupling processing module 1120, the first determination module 1130, the first prediction module 1140 and the generation module 1150 may be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application specific integrated circuit (ASIC), or may be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or by any one of the three implementation methods of software, hardware and firmware or by an appropriate combination of any of them. Alternatively, at least one of the first acquisition module 1110, the decoupling processing module 1120, the first determination module 1130, the first prediction module 1140 and the generation module 1150 may be at least partially implemented as a computer program module, and when the computer program module is run, the corresponding function may be performed.

[0210] It should be noted that the item recommendation device part in the embodiment of the present disclosure corresponds to the item recommendation method part in the embodiment of the present disclosure. The description of the item recommendation device part specifically refers to the item recommendation method part, which will not be repeated here.

[0211] Based on the above-mentioned training method of the recommendation model, the present disclosure also provides a training device for the recommendation model. Fig.11 The training device is described in detail.

[0212] Fig.11 A block diagram of a training device for a recommendation model according to an embodiment of the present disclosure is schematically shown.

[0213] like Fig.11 As shown, the training device 1200 of the recommendation model of this embodiment includes a preference decoupling module 1210 , a preference fusion module 1220 , a second prediction module 1230 and an iterative adjustment module 1240 .

[0214] The preference decoupling module 1210 is used to perform preference decoupling processing on the semantic features of sample items in each of the M sample domains to obtain sample domain shared features and sample domain specific features for each sample item, wherein the sample domain shared features represent features shared across domains, and the sample domain specific features represent features specific to the domain, and M ≥ 2. In one embodiment, the preference decoupling module 1210 can be used to perform the operation S410 described above, which will not be repeated here.

[0215] The preference fusion module 1220 is used to perform feature fusion processing according to the sample domain shared features and sample domain specific features of each sample item to obtain the sample fusion features of each sample item. In one embodiment, the preference fusion module 1220 can be used to perform the operation S420 described above, which will not be repeated here.

[0216] The second prediction module 1230 is used to predict the recommendation probability of each sample item according to the sample user characteristics and the sample fusion characteristics of each sample item, wherein the sample user characteristics represent the characteristics of users who interact with the sample items in the M sample fields. In one embodiment, the second prediction module 1230 can be used to perform the operation S430 described above, which will not be repeated here.

[0217] The iterative adjustment module 1240 is used to determine the recommendation loss information of the recommendation model based on the recommendation probability of each sample item, and iteratively adjust the network parameters of the recommendation model according to the recommendation loss information until the recommendation loss information satisfies the recommendation iteration stop condition, and the network parameters obtained when the recommendation iteration stop condition is satisfied are used as the trained recommendation model. In one embodiment, the iterative adjustment module 1240 can be used to perform the operation S440 described above, which will not be repeated here.

[0218] According to an embodiment of the present disclosure, the recommendation model includes a preference decoupling module, which includes a domain-shared encoder and a domain-specific encoder corresponding to each of the M sample domains.

[0219] According to an embodiment of the present disclosure, the preference decoupling module includes a second domain shared encoding submodule and a second domain specific encoding submodule;

[0220] The second domain-shared encoding submodule is used to input the semantic features of the sample items in the M sample domains into the domain-shared encoder, and output the sample domain-shared features of each sample item.

[0221] The second domain-specific encoding submodule is used to input the semantic features of the sample items in the M sample domains into the corresponding domain-specific encoders respectively, and output the sample domain-specific features of each sample item.

[0222] According to an embodiment of the present disclosure, the training device for the recommendation model further includes: a second determination module, a first adjustment module, a third determination module and a second adjustment module.

[0223] The second determination module is used to determine the specific feature similarity between every two sample domains using a first preset formula according to the sample domain specific features of each sample object.

[0224] The first adjustment module is used to adjust the network parameters of the domain-specific encoder using the specific feature similarity.

[0225] The third determination module is used to determine the similarity of shared features between every two sample domains using a second preset formula according to the shared features of the sample domains of each sample object.

[0226] The second adjustment module is used to adjust the network parameters of the domain shared encoder using the shared feature similarity.

[0227] According to an embodiment of the present disclosure, the recommendation model includes a preference fusion module.

[0228] The preference fusion module is used to perform feature fusion processing according to the sample domain shared features and sample domain specific features of each sample item to obtain the sample fusion features of each sample item.

[0229] According to an embodiment of the present disclosure, the preference fusion module includes: a third determination submodule and a fusion submodule.

[0230] The third determination submodule is used to determine the sample domain enhancement feature of each sample field according to the sample domain shared feature of each sample object.

[0231] The fusion submodule is used to perform feature fusion processing according to the sample domain shared features, sample domain specific features and sample domain enhanced features of each sample item, and output the sample fusion features of each sample item.

[0232] According to an embodiment of the present disclosure, the third determining submodule includes: a fusion unit.

[0233] The fusion unit is used to fuse the sample domain shared features of each sample item to obtain the sample domain enhanced features.

[0234] According to an embodiment of the present disclosure, the third determining submodule further includes: a clustering unit, a generating unit and a determining unit.

[0235] A clustering unit, used for clustering the sample domain shared features of each sample object to obtain at least one cluster;

[0236] A generation unit, used for generating a cluster-level cross-domain graph according to the cluster characteristics of each cluster, wherein the vertices of the cluster-level cross-domain graph correspond to the clusters, and the edges represent the similarities between clusters;

[0237] The determination unit is used to determine the sample domain enhancement feature of each sample domain according to the cluster-level cross-domain graph and each cluster cluster.

[0238] According to an embodiment of the present disclosure, the determining unit includes: an aggregation subunit and a determination subunit.

[0239] The aggregation subunit is used to perform domain sharing preference aggregation processing on the cluster-level cross-domain graph to obtain enhanced cluster features for each cluster, wherein the enhanced cluster feature representation adds inter-domain shared features to the cluster features;

[0240] The determination subunit is used to determine the sample domain enhanced features of each sample domain according to the cluster features and enhanced cluster features of each cluster cluster.

[0241] According to an embodiment of the present disclosure, the determining subunit includes: a first determining component and a second determining component.

[0242] A first determining component is used to determine, for a j-th sample field among the M sample fields, a j-th cluster allocation matrix of the j-th sample field according to the cluster characteristics of each cluster cluster, wherein the j-th cluster allocation matrix includes an allocation coefficient of each cluster cluster in the j-th sample field, 1≤j≤M;

[0243] The second determining component is used to determine the j-th sample domain enhanced feature of the j-th sample domain according to the j cluster allocation matrices and the enhanced cluster features of each cluster cluster.

[0244] According to an embodiment of the present disclosure, the second determination component includes: a multiplication subcomponent and a summation subcomponent.

[0245] A multiplication subcomponent, for each cluster, multiplies the enhanced cluster feature of the cluster by the distribution coefficient of the cluster in the j-th sample field to obtain the distribution feature of the cluster;

[0246] The summation subcomponent is used to sum the assigned features of each cluster to obtain the j-th sample domain enhanced features.

[0247] According to an embodiment of the present disclosure, the training device for the recommendation model further includes: a fourth determination module, a fifth determination module, a sixth determination module, a seventh determination module and an eighth determination module.

[0248] The fourth determination module is used to determine the sample item enhancement feature of each sample item according to the sample domain enhancement feature of each sample domain.

[0249] The fifth determination module is used to determine the enhanced domain shared features of the sample items according to the sample domain shared features and the sample item enhanced features of each sample item.

[0250] The sixth determination module is used to determine, for the i-th sample field among the M sample fields, a first preference score for the sample domain specific feature according to the sample user feature and the sample domain specific feature of the sample items in the i-th sample field.

[0251] The seventh determination module is used to determine a second preference score for the enhanced domain shared feature according to the sample user feature and the enhanced domain shared feature of the sample item.

[0252] An eighth determination module is used to determine fusion loss information for the preference fusion module based on the first preference score and the second preference score, and iteratively adjust the network parameters of the preference fusion module according to the fusion loss information until the fusion loss information meets the fusion iteration stop condition.

[0253] According to an embodiment of the present disclosure, the training device for the recommendation model further includes: a second acquisition module and a semantic extraction module.

[0254] The second acquisition module is used to acquire text data of sample items in the M sample fields and sample user information of users who interact with the sample items in the M sample fields.

[0255] The semantic extraction module is used to input the text data of the sample items and the sample user information into the semantic extraction module, and output the sample user features and the semantic features of the sample items of the M sample fields.

[0256] According to an embodiment of the present disclosure, any multiple of the preference decoupling module 1210, the preference fusion module 1220, the second prediction module 1230, and the iterative adjustment module 1240 can be combined in one module / unit / sub-unit for implementation, or any one of the modules / units / sub-units can be split into multiple modules / units / sub-units. Alternatively, at least part of the functions of one or more of these modules / units / sub-units can be combined with at least part of the functions of other modules / units / sub-units and implemented in one module / unit / sub-unit. According to an embodiment of the present disclosure, at least one of the preference decoupling module 1210, the preference fusion module 1220, the second prediction module 1230, and the iterative adjustment module 1240 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application specific integrated circuit (ASIC), or can be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or implemented in any one of the three implementation methods of software, hardware, and firmware, or in any appropriate combination of any of them. Alternatively, at least one of the preference decoupling module 1210 , the preference fusion module 1220 , the second prediction module 1230 , and the iterative adjustment module 1240 may be at least partially implemented as a computer program module, and when the computer program module is executed, a corresponding function may be performed.

[0257] It should be noted that the training device part of the recommended model in the embodiment of the present disclosure corresponds to the training method part of the recommended model in the embodiment of the present disclosure. The description of the training device part of the recommended model specifically refers to the training method part of the recommended model, which will not be repeated here.

[0258] Fig.12 A block diagram of an electronic device suitable for implementing the method described above according to an embodiment of the present disclosure is schematically shown. Fig.12 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0259] like Fig.12As shown, the electronic device 1300 according to an embodiment of the present disclosure includes a processor 1301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1302 or a program loaded from a storage part 1308 to a random access memory (RAM) 1303. The processor 1301 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or a related chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 1301 may also include an onboard memory for caching purposes. The processor 1301 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present disclosure.

[0260] In RAM 1303, various programs and data required for the operation of electronic device 1300 are stored. Processor 1301, ROM 1302 and RAM 1303 are connected to each other through bus 1304. Processor 1301 performs various operations of the method flow according to the embodiment of the present disclosure by executing the program in ROM 1302 and / or RAM 1303. It should be noted that the program can also be stored in one or more memories other than ROM 1302 and RAM 1303. Processor 1301 can also perform various operations of the method flow according to the embodiment of the present disclosure by executing the program stored in the one or more memories.

[0261] According to an embodiment of the present disclosure, the electronic device 1300 may further include an input / output (I / O) interface 1305, which is also connected to the bus 1304. The electronic device 1300 may further include one or more of the following components connected to the input / output (I / O) interface 1305: an input portion 1306 including a keyboard, a mouse, etc.; an output portion 1307 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage portion 1308 including a hard disk, etc.; and a communication portion 1309 including a network interface card such as a LAN card, a modem, etc. The communication portion 1309 performs communication processing via a network such as the Internet. The drive 1300 is also connected to the input / output (I / O) interface 1305 as needed. A removable medium 1301, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 1300 as needed, so that a computer program read therefrom is installed into the storage portion 1308 as needed.

[0262] According to an embodiment of the present disclosure, the method flow according to an embodiment of the present disclosure can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a computer-readable storage medium, and the computer program contains a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 1309, and / or installed from the removable medium 1301. When the computer program is executed by the processor 1301, the above-mentioned functions defined in the system of the embodiment of the present disclosure are executed. According to an embodiment of the present disclosure, the system, equipment, device, module, unit, etc. described above can be implemented by a computer program module.

[0263] The present disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or may exist independently without being assembled into the device / apparatus / system. The above computer-readable storage medium carries one or more programs, and when the above one or more programs are executed, the method according to the embodiment of the present disclosure is implemented.

[0264] According to an embodiment of the present disclosure, the computer-readable storage medium may be a non-volatile computer-readable storage medium. For example, it may include, but is not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in combination with an instruction execution system, apparatus, or device.

[0265] For example, according to an embodiment of the present disclosure, the computer-readable storage medium may include the ROM 1302 and / or the RAM 1303 described above and / or one or more memories other than the ROM 1302 and the RAM 1303 .

[0266] The embodiments of the present disclosure also include a computer program product, which includes a computer program, and the computer program contains program code for executing the method provided by the embodiments of the present disclosure. When the computer program product runs on an electronic device, the program code is used to enable the electronic device to implement the above method provided by the embodiments of the present disclosure.

[0267] When the computer program is executed by the processor 1301, the above functions defined in the system / device of the embodiment of the present disclosure are executed. According to the embodiment of the present disclosure, the system, device, module, unit, etc. described above can be implemented by a computer program module.

[0268] In one embodiment, the computer program may be based on a tangible storage medium such as an optical storage device, a magnetic storage device, etc. In another embodiment, the computer program may also be transmitted and distributed in the form of a signal on a network medium, and downloaded and installed through the communication part 1309, and / or installed from the removable medium 1301. The program code contained in the computer program may be transmitted using any appropriate network medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.

[0269] According to an embodiment of the present disclosure, the program code for executing the computer program provided by the embodiment of the present disclosure can be written in any combination of one or more programming languages. Specifically, these computing programs can be implemented using high-level process and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, Java, C++, python, "C" language or similar programming languages. The program code can be executed entirely on the user computing device, partially on the user device, partially on the remote computing device, or entirely on the remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (for example, using an Internet service provider to connect through the Internet).

[0270] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram may represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box may also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions. It can be understood by those skilled in the art that the features recorded in the various embodiments of the present disclosure can be combined and / or combined in a variety of ways, even if such a combination or combination is not explicitly recorded in the present disclosure. In particular, without departing from the spirit and teaching of the present disclosure, the features described in the various embodiments of the present disclosure may be combined and / or combined in a variety of ways. All of these combinations and / or combinations fall within the scope of the present disclosure.

[0271] The embodiments of the present disclosure are described above. However, these embodiments are only for illustrative purposes and are not intended to limit the scope of the present disclosure. Although the embodiments are described above, this does not mean that the measures in the various embodiments cannot be used in combination to advantage. Without departing from the scope of the present disclosure, those skilled in the art may make a variety of substitutions and modifications, which should all fall within the scope of the present disclosure.

Claims

1. An item recommendation method, comprising: In response to an operation of a target user in a target application, acquiring target user information of the target user and item information of at least one item to be recommended in the target application; Performing preference decoupling processing on item information of at least one of the items to be recommended, to obtain target domain shared features and target domain specific features of each of the items to be recommended; According to the target domain to which the target application belongs, determining a target domain enhancement feature corresponding to the target domain; Predicting the recommendation probability of each of the recommended items according to the target user information, the target domain shared features, the target domain specific features and the target domain enhanced features of each of the items to be recommended; as well as Generate an item recommendation list for the target user according to the recommendation probability of at least one of the items to be recommended.

2. The method according to claim 1, wherein: The performing preference decoupling processing on the item information of at least one of the items to be recommended to obtain the target domain shared features and target domain specific features of each of the items to be recommended includes: Performing domain sharing encoding processing on item information of at least one of the items to be recommended to obtain target domain sharing features of each of the recommended items; Domain-specific encoding is performed on item information of at least one of the items to be recommended to obtain target domain-specific features of each of the recommended items.

3. The method according to claim 1, wherein: The determining, according to the target domain to which the target application belongs, the target domain enhancement feature corresponding to the target domain comprises: According to the target domain to which the target application belongs, a target domain enhancement feature for the target domain is determined from a domain-feature relationship graph, wherein the domain-feature relationship graph includes at least one domain and a domain enhancement feature corresponding to each of the domains.

4. The method according to claim 1, wherein: The predicting the recommendation probability of each of the recommended items according to the target user information, the target domain shared features, the target domain specific features and the target domain enhanced features of each of the items to be recommended comprises: For each of the items to be recommended, determining a target fusion feature of each of the items to be recommended according to the target domain shared feature, the target domain specific feature and the target domain enhanced feature; The recommendation probability of each of the recommended items is predicted according to the target user information and the target fusion feature of each of the items to be recommended.

5. The method according to claim 1, further comprising: Extracting semantic features from the target user information and at least one of the item information to be recommended, to obtain target user features and item features to be recommended of each of the items to be recommended; The performing preference decoupling processing on the item information of at least one of the items to be recommended to obtain the target domain shared features and target domain specific features of each of the items to be recommended includes: Performing preference decoupling processing on the features of each of the items to be recommended, so as to obtain the target domain shared features and target domain specific features of each of the items to be recommended; The predicting the recommendation probability of each of the recommended items according to the target user information, the target domain shared features, the target domain specific features and the target domain enhanced features of each of the to-be-recommended items comprises: The recommendation probability of each of the recommended items is predicted according to the target user features, the target domain shared features, the target domain specific features and the target domain enhanced features of each of the to-be-recommended items.

6. A training method for a recommendation model, comprising: Perform preference decoupling processing on the semantic features of sample items in each of the M sample domains to obtain sample domain shared features and sample domain specific features of each of the sample items, wherein the sample domain shared features represent features shared across domains, and the sample domain specific features represent features specific to a domain, M≥2; Perform feature fusion processing according to the sample domain shared features and the sample domain specific features of each of the sample items to obtain a sample fusion feature of each of the sample items; Predicting the recommendation probability of each of the sample items according to sample user features and sample fusion features of each of the sample items, wherein the sample user features represent features of users who interact with the sample items in the M sample fields; and The recommendation loss information of the recommendation model is determined based on the recommendation probability of each of the sample items, and the network parameters of the recommendation model are iteratively adjusted according to the recommendation loss information until the recommendation loss information meets the recommendation iteration stopping condition, and the network parameters obtained when the recommendation iteration stopping condition is met are used as the trained recommendation model.

7. The method according to claim 6, wherein: The recommendation model includes a preference decoupling module, wherein the preference decoupling module includes a domain-shared encoder and a domain-specific encoder corresponding to each of the M sample domains; The preference decoupling process is performed on the semantic features of the sample items in the M sample domains to obtain the sample domain shared features and sample domain specific features of each sample item, including: Inputting the semantic features of sample items of the M sample domains into the domain-sharing encoder, and outputting the sample domain-sharing features of each sample item; The semantic features of the sample items in each of the M sample domains are respectively input into the corresponding domain-specific encoders, and the sample domain-specific features of each of the sample items are output.

8. The method according to claim 7, further comprising: Determine the similarity of the specific features between each two sample domains using a first preset formula according to the sample domain specific features of each of the sample objects; adjusting network parameters of the domain-specific encoder using the specific feature similarity; Determine the similarity of the shared features between every two sample domains using a second preset formula according to the shared features of the sample domains of each of the sample objects; The shared feature similarity is used to adjust network parameters of the domain-shared encoder.

9. The method according to claim 6, wherein: The recommendation model includes a preference fusion module; The performing feature fusion processing according to the sample domain shared features and the sample domain specific features of each of the sample items to obtain the sample fusion features of each of the sample items comprises: Inputting the sample domain shared features and the sample domain specific features of each of the sample items into the preference fusion module, and determining the sample domain enhanced features of each of the sample domains according to the sample domain shared features of each of the sample items; Feature fusion processing is performed according to the sample domain shared features, the sample domain specific features and the sample domain enhanced features of each of the sample items, and the sample fusion features of each of the sample items are output.

10. The method according to claim 9, wherein: Determining the sample domain enhancement feature of each sample field according to the sample domain shared feature of each sample object comprises: Performing fusion processing on the sample domain shared features of each of the sample items to obtain the sample domain enhanced features; or Performing clustering processing on the sample domain shared features of each of the sample objects to obtain at least one cluster; Generate a cluster-level cross-domain graph according to the cluster features of each cluster, wherein the vertices of the cluster-level cross-domain graph correspond to the clusters and the edges represent the similarities between clusters; A sample domain enhancement feature of each of the sample domains is determined according to the cluster-level cross-domain graph and each of the clustering clusters.

11. The method according to claim 10, wherein: Determining the sample domain enhancement feature of each sample domain according to the cluster-level cross-domain graph and each of the cluster clusters includes: Performing domain sharing preference aggregation processing on the cluster-level cross-domain graph to obtain enhanced cluster features of each cluster, wherein the enhanced cluster features represent the addition of inter-domain sharing features to the cluster features; The sample domain enhanced feature of each sample domain is determined according to the cluster feature and the enhanced cluster feature of each cluster cluster.

12. The method according to claim 11, wherein: Determining the sample domain enhanced feature of each sample domain according to the cluster feature and the enhanced cluster feature of each cluster cluster includes: For a j-th sample field among the M sample fields, determining a j-th cluster allocation matrix of the j-th sample field according to the cluster features of each of the cluster clusters, wherein the j-th cluster allocation matrix includes allocation coefficients of each of the cluster clusters in the j-th sample field, 1≤j≤M; The j-th sample domain enhanced feature of the j-th sample domain is determined according to the j cluster allocation matrices and the enhanced cluster feature of each of the cluster clusters.

13. The method according to claim 12, wherein: The determining the jth sample domain enhanced feature of the jth sample domain according to the j cluster allocation matrices and the enhanced cluster features of each of the cluster clusters comprises: For each of the clusters, multiplying the enhanced cluster feature of the cluster by the distribution coefficient of the cluster in the j-th sample field to obtain the distribution feature of the cluster; The j-th sample domain enhancement feature is obtained by summing up the assigned features of each of the clusters.

14. The method according to claim 9, further comprising: Determine the sample item enhancement feature of each of the sample items according to the sample domain enhancement feature of each of the sample domains; Determine the enhanced domain shared features of the sample items according to the sample domain shared features and the sample item enhanced features of each of the sample items; For an i-th sample field among the M sample fields, determining a first preference score for the sample domain specific feature according to the sample user feature and the sample domain specific feature of the sample item in the i-th sample field; determining a second preference score for the enhanced domain shared feature according to the sample user feature and the enhanced domain shared feature of the sample item; Based on the first preference score and the second preference score, fusion loss information for the preference fusion module is determined, and network parameters of the preference fusion module are iteratively adjusted according to the fusion loss information until the fusion loss information meets a fusion iteration stop condition.

15. The method according to claim 6, wherein: The recommendation model also includes a semantic extraction module; The method further comprises: Acquire text data of sample items in the M sample fields and sample user information of users who interact with the sample items in the M sample fields; The text data of the sample items and the sample user information are input into the semantic extraction module, and the sample user features and the semantic features of the sample items of the M sample fields are output.

16. An item recommendation device, comprising: A first acquisition module, configured to acquire target user information of the target user and item information of at least one item to be recommended in the target application in response to an operation of the target user in the target application; a decoupling processing module, configured to perform preference decoupling processing on the item information of at least one of the items to be recommended, so as to obtain target domain shared features and target domain specific features of each of the items to be recommended; A first determination module is used to determine a target domain enhancement feature corresponding to the target domain according to the target domain to which the target application belongs; A first prediction module, configured to predict the recommendation probability of each of the recommended items according to the target user information, the target domain shared features, the target domain specific features and the target domain enhanced features of each of the items to be recommended; as well as A generating module is used to generate an item recommendation list for the target user according to the recommendation probability of at least one of the items to be recommended.

17. A training device for a recommendation model, comprising: A preference decoupling module is used to perform preference decoupling processing on the semantic features of sample items in each of the M sample domains to obtain sample domain shared features and sample domain specific features of each of the sample items, wherein the sample domain shared features represent features shared across domains, and the sample domain specific features represent features specific to a domain, M≥2; A preference fusion module, used for performing feature fusion processing according to the sample domain shared features and the sample domain specific features of each of the sample items, to obtain a sample fusion feature of each of the sample items; A second prediction module is configured to predict the recommendation probability of each of the sample items according to sample user characteristics and sample fusion characteristics of each of the sample items, wherein the sample user characteristics represent characteristics of users who interact with the sample items in the M sample fields; and An iterative adjustment module is used to determine the recommendation loss information of the recommendation model based on the recommendation probability of each of the sample items, and iteratively adjust the network parameters of the recommendation model according to the recommendation loss information until the recommendation loss information meets the recommendation iteration stopping condition, and the network parameters obtained when the recommendation iteration stopping condition is met are used as the trained recommendation model.

18. An electronic device, comprising: one or more processors; a memory for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 15.

19. A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the method according to any one of claims 1 to 15.

20. A computer program product comprising a computer program which, when executed by a processor, implements the method according to any one of claims 1 to 15.

Images