Object recommendation method and apparatus

Abstract

Embodiments of the present specification provide an object recommendation method and device, wherein the object recommendation method comprises: obtaining user information of a target user, domain-in interaction information and cross-domain interaction information learned in advance, wherein the domain-in interaction information represents interaction between the target user and an interaction object within a data domain, and the cross-domain interaction information represents indirect interaction between the target user and the interaction object among multiple data domains; determining a target interaction object based on the user information, the domain-in interaction information and the cross-domain interaction information; and recommending the target interaction object to the target user. In the present scheme, not only the domain-in interaction information between the target user and the interaction object is considered, but also higher-order and deeper cross-domain interaction information between the target user and the interaction object is considered when making object recommendation, so that the determined target interaction object is more in line with the preferences of the target user, and the accuracy of object recommendation is improved.

Description

TECHNICAL FIELD

[0001] Embodiments of the present specification relate to the technical field of data processing, in particular to an object recommendation method. BACKGROUND

[0002] With the rapid development of the Internet, information resources on the network grow rapidly, and it is very difficult for users to find information meeting their needs from a large amount of text, video, image and commodity resources. In order to improve user experience, a recommendation system capable of personalized recommendation for users is proposed, which analyzes the interest preferences of users by studying the historical behavior data of users, and then recommends resources meeting the needs of users to users.

[0003] However, many projects currently encounter the problem of data sparsity, which affects the recommendation effect of many recommendation systems. In order to solve this problem, a cross-domain recommendation method (CDR, Cross-domain Recommendation) is proposed, that is, for the problem of data sparsity in a certain data domain, historical behavior data of users in other data domains can be used to recommend objects in the data domain to users. However, since this method involves cross-domain, the accuracy is relatively low compared with the method of domain recommendation. Therefore, there is an urgent need for a method capable of accurately recommending objects. SUMMARY

[0004] Therefore, one or more embodiments of the present specification provide an object recommendation method. One or more embodiments of the present specification also provide an object recommendation device, a computing device, a computer-readable storage medium and a computer program to solve the technical defects in the prior art.

[0005] According to a first aspect of an embodiment of the present specification, an object recommendation method is provided, comprising:

[0006] obtaining user information of a target user, pre-learned intra-domain interaction information and cross-domain interaction information, wherein the intra-domain interaction information represents the interaction between the user and the interaction object within the data domain, and the cross-domain interaction information represents the indirect interaction between the user and the interaction object among multiple data domains;

[0007] determining a target interaction object based on the user information, the intra-domain interaction information and the cross-domain interaction information;

[0008] recommending the target interaction object to the target user.

[0009] According to a second aspect of an embodiment of the present specification, an object recommendation device is provided, comprising:

[0010] An obtaining module configured to obtain user information of a target user, domain-intra interaction information and cross-domain interaction information obtained through pre-learning, wherein the domain-intra interaction information represents interaction between the target user and an interaction object within a data domain, and the cross-domain interaction information represents indirect interaction between the target user and the interaction object among a plurality of data domains;

[0011] A determining module configured to determine a target interaction object based on the user information, the domain-intra interaction information and the cross-domain interaction information.

[0012] A recommending module configured to recommend the target interaction object to the target user.

[0013] According to a third aspect of the embodiments of the present specification, a computing device is provided, comprising:

[0014] a memory and a processor;

[0015] The memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions, and the computer executable instructions, when executed by the processor, implement the steps of the object recommendation method.

[0016] According to a fourth aspect of the embodiments of the present specification, a computer readable storage medium is provided, which stores computer executable instructions, and the instructions, when executed by a processor, implement the steps of the object recommendation method.

[0017] According to a fifth aspect of the embodiments of the present specification, a computer program is provided, and when the computer program is executed in a computer, the computer program causes the computer to execute the steps of the object recommendation method.

[0018] An embodiment of the specification realizes obtaining user information of a target user, domain-intra interaction information learned in advance, and cross-domain interaction information, wherein the domain-intra interaction information represents interaction between the user and an interaction object within a data domain, and the cross-domain interaction information represents indirect interaction between the user and the interaction object among multiple data domains; determining a target interaction object based on the user information, the domain-intra interaction information, and the cross-domain interaction information; and recommending the target interaction object to the target user. That is, in the present scheme, the domain-intra interaction information between the user and the interaction object within the data domain, and the cross-domain interaction information between the user and the interaction object among different data domains are learned in advance. Since the user and the interaction object have interaction within the domain, it indicates that the user is interested in the interaction object, and the user and the interaction object have indirect interaction relationship among different data domains, which indicates that the user and the interaction object may not have interaction at present, but according to the cross-domain association between users, the cross-domain association between interaction objects, and the domain-intra interaction relationship between the user and the interaction object, it can be determined that the user and the interaction object are likely to have interaction. Therefore, the interaction object can be recommended to the user when the object is recommended. The present scheme extracts higher-order and deeper cross-domain interaction information between the user and the interaction object, and considers more abundant interaction information when the object is recommended. Therefore, the target interaction object determined is more in line with the preference of the target user, improves the accuracy of object recommendation, and in the case of sparse data in a certain data domain, since the indirect interaction relationship between the user and the interaction object among different data domains is extracted, accurate object recommendation can be realized for the user in the data domain based on the indirect interaction relationship. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A processing flowchart of an object recommendation method executed in an object recommendation system is shown according to an embodiment of the specification;

[0020] Figure 2 A processing flowchart of another object recommendation method executed in an object recommendation system is shown according to an embodiment of the specification;

[0021] Figure 3 A flowchart of an object recommendation method is shown according to an embodiment of the specification;

[0022] Figure 4 A page schematic diagram of a target interaction object display is shown according to an embodiment of the specification;

[0023] Figure 5 A flowchart of a recommendation model training method is shown according to an embodiment of the specification;

[0024] Figure 6 A schematic diagram of a domain-intra interaction graph is shown according to an embodiment of the specification;

[0025] Figure 7 A schematic diagram of a cross-domain interaction graph is shown according to an embodiment of the present specification;

[0026] Figure 8 A schematic diagram of another cross-domain interaction graph is shown according to an embodiment of the present specification;

[0027] Figure 9 A schematic diagram of yet another cross-domain interaction graph is shown according to an embodiment of the present specification;

[0028] Figure 10 A schematic diagram of determining a correlation feature vector based on a cross-domain interaction graph is shown according to an embodiment of the present specification;

[0029] Figure 11 A flowchart of a processing procedure of a feature vector is shown according to an embodiment of the present specification;

[0030] Figure 12 A flowchart of a recommendation model training method is shown according to an embodiment of the present specification;

[0031] Figure 13 A flowchart of a processing procedure of an object recommendation method applied to a traffic marketing scenario is shown according to an embodiment of the present specification;

[0032] Figure 14 A structural schematic diagram of an object recommendation apparatus is shown according to an embodiment of the present specification;

[0033] Figure 15 A structural block diagram of a computing device is shown according to an embodiment of the present specification. DETAILED DESCRIPTION

[0034] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present specification. However, the present specification can be practiced without the specific details, which are not described in the present specification, and it is understood that the scope of the present specification is not limited to the details below. In other instances, well-known methods associated with computing, software development, and the like have not been described in detail in order to avoid unnecessarily obscuring aspects of the present specification.

[0035] The terminology used in one or more embodiments of the present specification is for the purpose of describing particular embodiments only and is not intended to be limiting of one or more embodiments of the present specification. As used in one or more embodiments of the present specification and the accompanying claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in one or more embodiments of the present specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0036] It should be understood that, although the terms first, second, etc. can be employed in describing various information in one or more embodiments of the present specification, the information should not be limited to such terms. These terms are only used to distinguish one category of information from another category of information. For example, without departing from the scope of one or more embodiments of the present specification, first can also be referred to as second, and similarly, second can also be referred to as first. Depending on the context, the word "if' as used herein can be interpreted as meaning "when" or "upon" or "in response to determining."

[0037] First, the noun terms related to one or more embodiments of the present specification are explained.

[0038] Graph: A graph is a structure that describes a set of objects in which certain pairs of objects are "related" in some sense. These pairs of objects correspond to mathematical abstractions called vertices (also called nodes or points), and each related pair of vertices is called an edge (also called a link or a line).

[0039] Graph Neural Network: A neural network that can process graph-structured data, which updates the features on the target node by aggregating or passing information in the neighborhood of the target node, and then performs downstream tasks.

[0040] Hypergraph: A graph in which each edge contains more than two nodes, which has the ability to describe high-order relationships between nodes compared to ordinary graphs.

[0041] Hypergraph Neural Network: A neural network that can process hypergraph-structured data and learn high-order relationships between nodes.

[0042] Embedding: A layer commonly used in deep learning network models, mainly used to process vector representations of sparse features. It not only solves the length problem of one-hot vectors, but also represents the similarity between features.

[0043] Domain: In the field of machine learning, a sample set with the same data distribution is usually called a "domain". The concept of domain in a recommendation system is mainly divided according to different project scenarios, such as "book recommendation" and "movie recommendation", which belong to two different domains.

[0044] Hyperbolic space: A manifold space with negative constant curvature, which has the ability to express hierarchical structure, for example, data itself has a certain hierarchical structure, which can be restored in hyperbolic space.

[0045] Intra-domain interaction information: represents the interaction relationship between users and interaction objects within the same data domain. For example, in data domain 1, user A has an interaction relationship with interaction object z, and this relationship can be intra-domain interaction information.

[0046] Cross-domain interaction information: represents the indirect interaction relationship between users and interaction objects in different data domains. For example, in data domain 1, user A and user B both have interaction relationships with interaction object z, and in data domain 2, user A has an interaction relationship with interaction object m, but user B has no historical data in the data domain. It can be determined that there is an indirect interaction relationship between user B and interaction object m in data domain 2, and this relationship can be cross-domain interaction information.

[0047] Machine learning model: an algorithm expression that can find patterns or make predictions by analyzing large amounts of data. Machine learning models are powered by data and are the mathematical engine of artificial intelligence.

[0048] Intra-domain interaction graph: a heterogeneous graph constructed based on the interaction relationship between users and interaction objects in the same data domain.

[0049] Cross-domain interaction graph: a hypergraph constructed based on the association relationship between users and users in different data domains, or a hypergraph constructed based on the association relationship between interaction objects and interaction objects in different data domains.

[0050] Interaction label: used to represent whether there is an interaction relationship between users and interaction objects, and is a label in the training of recommendation models.

[0051] Feature extraction layer: a layer used for linear mapping, nonlinear transformation, and feature extraction of input vectors.

[0052] Fully connected layer: used to map all previously extracted feature vectors (high-dimensional feature maps) into one-dimensional feature vectors, which contain all feature information and can be converted into predicted interaction indicators.

[0053] One-hot encoding: also known as one-hot encoding, is a feature extraction method that uses N-bit state registers to encode N states, with each state having an independent register bit, and only one bit in these register bits is valid.

[0054] With the popularization of intelligent recommendation, many projects have encountered the problem of data sparsity, that is, a small number of users contribute to most user-item interactions. This problem affects the recommendation effect of long-tail users / items and new users / items (cold start) in many recommendation systems.

[0055] To solve this problem, cross-domain recommendation method (CDR) is proposed by scholars. CDR can migrate the knowledge of the information-rich domain to the data-sparse domain to help it improve the effect. CDR problems are often divided into three categories according to the target, STCDR (Single-Target Cross-Domain Recommendation), DTCDR (Dual-Target Cross-Domain Recommendation) and MTCDR (Multi-Target Cross-Domain Recommendation). STCDR is to use the information-rich A domain to improve the effect of the data-sparse B domain, DTCDR is to improve the effect of A domain and B domain at the same time. MTCDR is more complex than them, which is to improve the effect of multiple domains (3 or more) at the same time.

[0056] The core problem of CDR is how to realize the migration of inter-domain information in a better way. In many studies, embedding migration or content migration method is used for inter-domain information transmission of STCDR and DTCDR. This method is usually based on common users or common items between domains, using embedding or attribute feature mapping to realize transfer learning. This approach performs well in STCDR and DTCDR. But if such ideas are directly applied to MTCDR, it will be a disaster, because they are basically pair-wise (pair-wise combination) approach, if used in MTCDR, the complexity of constructing pair-wise will be N^2 complexity, which is relatively difficult to implement.

[0057] In addition, the classic graph scheme for cross-domain recommendation is the Hero Graph framework, which first introduced the graph method into MTCDR. They constructed an inter-domain shared heterogeneous graph with multiple domains of users and items as different nodes, and if a user bought an item, they formed an edge. Then each domain is modeled separately, and the shared heterogeneous graph is used to handle cross-domain information. Hero Graph provides a good solution for MTCDR, but it also has two shortcomings. One is that it does not consider higher-order relationships in user-item interaction data, because the complexity of the recommendation system in the real world is high, and the two-relationship modeling of ordinary graph methods is already difficult to represent; the second is that it ignores the hierarchical relationship between users and items, such as goods, which usually have popularity (hotness), category structure (category) and other hierarchical relationships, which are difficult to capture through graph embedding.

[0058] To this end, the present specification provides an object recommendation method capable of solving the above problems, and specific implementations can be referred to the relevant descriptions of each of the following embodiments.

[0059] In the present specification, an object recommendation method is provided, and the present specification also relates to an object recommendation device, a computing device, and a computer-readable storage medium, which are described in detail one by one in the following embodiments.

[0060] Referring to Figure 1 , Figure 1 A processing flowchart of an object recommendation method executed in an object recommendation system according to an embodiment of the present specification is shown.

[0061] As Figure 1 shown, the object recommendation system includes a server 101 and a client 102. The server 101 can provide background services for the client 102. A user registers or logs in the client 102, and then the client 102 can obtain user information of the user according to the user operation. The server 101 can obtain the user information of the target user from the client 102, and then the intra-domain interaction information and the cross-domain interaction information are pre-learned and can be considered to be stored in the server 101. Therefore, the server 101 can obtain the intra-domain interaction information and the cross-domain interaction information from the local. The intra-domain interaction information represents the interaction between the user and the interaction object within the data domain, and the cross-domain interaction information represents the indirect interaction between the user and the interaction object among multiple data domains. Then, based on the user information, the intra-domain interaction information and the cross-domain interaction information, the target interaction object is determined, and then the target interaction object is sent to the client 102 of the target user, i.e., the target interaction object is recommended to the target user. After the client 102 receives the target interaction object, the target interaction object is displayed, and then the user can see the recommended target interaction object.

[0062] The object recommendation method provided by the embodiments of the present specification pre-learns the intra-domain interaction information of users and interaction objects in a data domain, and the cross-domain interaction information of users and interaction objects between different data domains. Since the user and the interaction object have interaction in the domain, it indicates that the user is very interested in the interaction object, and the interaction object can be recommended to the user. There is an indirect interaction relationship between the user and the interaction object in different data domains, which indicates that the user and the interaction object may not have interaction at present, but according to the cross-domain association between users, the cross-domain association between interaction objects, and the intra-domain interaction relationship between users and interaction objects, it can be determined that the user and the interaction object are likely to have interaction. Therefore, when recommending objects to the user, the interaction object can be recommended. That is, the present scheme extracts higher-order and deeper cross-domain interaction information between users and interaction objects, and considers more abundant interaction information when making object recommendation. Therefore, the determined target interaction object is more in line with the preferences of the target user, and the accuracy of object recommendation is improved. Moreover, in the case of sparse data in a certain data domain, since the indirect interaction relationship between users and interaction objects in different data domains is extracted, accurate object recommendation can be realized for the user in the data domain based on the indirect interaction relationship.

[0063] Referring to Figure 2 , Figure 2 A processing flow chart of another object recommendation method executed in an object recommendation system according to an embodiment of the present specification is shown.

[0064] As Figure 2 shown, the object recommendation system includes a server 101 and a client 102, and the server 101 includes a training unit 1011 and an application unit 1012.

[0065] The training unit 1011 of the server obtains a plurality of sample groups of at least two data domains and corresponding interaction labels of each sample group from the client 102, constructs an intra-domain interaction graph corresponding to each data domain and a cross-domain interaction graph between different data domains based on the plurality of sample groups, determines the interaction feature vector of the user and the interaction feature vector of the interaction object in the intra-domain interaction graph by using the graph neural network, determines the correlation feature vector of each node in the cross-domain interaction graph by using the hypergraph neural network, and determines the predicted interaction index corresponding to each sample group according to the correlation feature vector and the interaction feature vector, wherein the predicted interaction index is used to represent the probability of interaction between the user and the interaction object. Then, the parameters of the recommendation model are adjusted based on the predicted interaction index corresponding to each sample group and the interaction label corresponding to each sample group until the training stopping condition is reached, and the trained recommendation model is obtained. Since the calculation process of the interaction feature vector is based on the aggregation of the feature information of the node and its neighbor nodes, it can be considered that the graph neural network model learns the intra-domain interaction information between the user and the interaction object in the process of determining the interaction feature vector. Since the correlation feature vector is determined according to the feature information of the reference node and the feature information of other nodes corresponding to the same hyperedge of the reference node, and the nodes in the hypergraph are cross-domain correlated, it can be considered that the hypergraph neural network learns the cross-domain interaction information between the nodes in the process of determining the correlation feature vector of the node.

[0066] Then the training unit 1011 of the server sends the trained recommendation model to the application unit 1012 of the server. The application unit 1012 of the server obtains the user information of the target user from the client 102, inputs the user information into the recommendation model, and determines the target interaction object by using the intra-domain interaction information and the cross-domain interaction information in the recommendation model. Then the application unit 1012 of the server sends the target interaction object to the client 102, and the client 102 displays the target interaction object for the target user.

[0067] The object recommendation method provided by the embodiments of the present specification first trains to obtain a recommendation model, that is, according to a sample group and an interaction label corresponding to the sample group, determines an association feature vector, an interaction feature vector of a user, and an interaction feature vector of an interaction object, and learns cross-domain interaction information in the process of determining the association feature vector and learns in-domain interaction information in the process of determining the interaction feature vector, and then determines a predicted interaction index between the user and the interaction object in each sample group according to the association feature vector and the interaction feature vector, adjusts the parameters of the recommendation model according to the predicted interaction index and the interaction label, and stops until the training condition is met, and the training of the recommendation model is completed. The user information obtained is input into the recommendation model, and the user information is processed according to the in-domain interaction information and the cross-domain interaction information to determine the target interaction object recommended for the target user. That is, when the object is recommended, not only the interaction object that has actually interacted with the user is considered, but also the interaction object that may interact with the user is determined according to the cross-domain interaction information. The target interaction object recommended based on the more comprehensive information will be more in line with the interest of the target user, that is, the accuracy of the result of the object recommendation is higher.

[0068] Referring to Figure 3 , Figure 3 A flowchart of an object recommendation method according to an embodiment of the present specification is shown, which can specifically include the following steps.

[0069] Step 302: obtaining user information of a target user, pre-learned in-domain interaction information, and cross-domain interaction information, wherein the in-domain interaction information represents the interaction between the user and the interaction object within the data domain, and the cross-domain interaction information represents the indirect interaction between the user and the interaction object between multiple data domains.

[0070] The target user can be a user waiting for the client to recommend an interaction object to him / her. The user information can include information that can uniquely identify a user, for example, the user information can be a user identifier, such as a user account, a user nickname, a user ID, etc. The user information can also include information that can represent the user's interest or hobby, for example, the user information can be the user's attribute information, such as the user's basic information, the user's historical behavior data, etc., wherein the user's historical behavior data can include the user's likes, collections, comments, forwards, attentions, shares, and browsing, etc.

[0071] The interaction object can be various information displayed to the user in the data domain. For example, the interaction object can be an advertisement, a text, a link, a picture, a video, a commodity, etc.

[0072] The data domain can be a domain composed of data corresponding to different projects according to different project scenarios. For example, a shopping scenario, a video playing scenario, and a music playing scenario correspond to different data domains, respectively.

[0073] As an example, the interaction of the user with the interactive object can be clicking, browsing, playing, converting, charging, purchasing, collecting, commenting, and the like of the user on the interactive object. For example, if user A accesses interactive object a in data field 1, it can be determined that there is an interaction relationship between user A and interactive object a, which is the intra-field interaction information.

[0074] As an example, indirect interaction refers to no interaction at present but predicted possible interaction, and the indirect interaction of the user with the interactive object refers to that the user has no interaction with the interactive object at present, but according to the interaction of the user with other interactive objects and / or the interaction of other users with the interactive object, it is determined that the user and the interactive object can have interaction. That is, the indirect interaction of the user with the interactive object between different data fields can be understood as that the user and the interactive object have no interaction, but according to the interaction relationship between other users associated with the user and the interactive object, and / or according to the interaction relationship between other interactive objects associated with the interactive object and the user, it can be determined that the user and the interactive object can have interaction. For example, user A browses interactive object a in data field 1 and purchases interactive object b in data field 2, and user B browses interactive object a in data field 1, so it can be determined that user B has an indirect interaction relationship with interactive object b in data field 2, and user B can also purchase interactive object b in data field 2, which is the cross-field interaction information.

[0075] In one or more embodiments of the present specification, since some data fields have the problem of data sparsity, the interaction relationship between the user and the interactive object in the data field cannot accurately recommend the interactive object to the user; or some users use the software corresponding to the data field for the first time (i.e., the cold start problem), and there is no historical behavior data of the user in the data field for reference, and it is difficult to achieve accurate recommendation for the user. Therefore, the present scheme pre-learns the cross-field interaction information, i.e., the indirect interaction relationship between the user and the interactive object in multiple data fields, so as to achieve the effect of recommending the interactive object to the user in the data field according to the data of other data fields.

[0076] In some embodiments, the user information of the target user can be obtained from the client. The user needs to register or log in to use the client, and the client can obtain the user information of the user according to the information input by the user and send the user information to the server, so that the server can obtain the user information of the target user. The pre-learned intra-field interaction information and cross-field interaction information can be learned by processing the historical interaction data of the user and the interactive object in multiple data fields, and can be stored in the local cache of the server, so that the server can obtain the pre-learned intra-field interaction information and cross-field interaction information from the local cache.

[0077] In the embodiments of the present application, since the user information of the target user, the pre-learned domain interaction information and the cross-domain interaction information are needed when determining the target interaction object, the information is first obtained in this step, thereby providing a data basis for subsequent determination of the target interaction object.

[0078] Step 304: determining the target interaction object based on the user information, the domain interaction information and the cross-domain interaction information.

[0079] The target interaction object is an interaction object that is more in line with the preferences of the target user and is an interaction object that the target user is likely to interact with.

[0080] In a possible implementation of the present application, the domain interaction information and the cross-domain interaction information are obtained by pre-training a recommendation model, and the recommendation model is a machine learning model.

[0081] The specific implementation of determining the target interaction object based on the user information, the domain interaction information and the cross-domain interaction information can include:

[0082] The user information is input into the recommendation model, and the recommendation model is used to operate the user information by using the domain interaction information and the cross-domain interaction information, thereby obtaining the target interaction object.

[0083] The recommendation model is a machine learning model used to recommend the target interaction object for the user, and can be pre-trained.

[0084] That is, the domain interaction information and the cross-domain interaction information can be pre-learned by training the recommendation model, that is, in the present application, the target interaction object can be recommended for the target user by the recommendation model.

[0085] In some embodiments of the present application, the recommendation model is pre-trained so that the recommendation model can learn the domain interaction information and the cross-domain interaction information. It can also be understood that the domain interaction information and the cross-domain interaction information are used as parameters of the recommendation model. After the user information is input into the recommendation model, the user information can be processed by using the domain interaction information and the cross-domain interaction information, and the recommendation model can output the target interaction object.

[0086] As an example, since the recommendation model is used to recommend the interaction object to the user, the interaction object information of the interaction object is also input when the recommendation model is pre-trained, so that the recommendation model can directly recommend the target interaction object to the target user, and the recommendation model can also learn and store the feature vector of the interaction object during the training process. After the user information is input into the recommendation model, the user information can be processed according to the domain interaction information and the cross-domain interaction information to obtain the enhanced feature vector of the target user, and then the enhanced feature vector of the target user and the enhanced feature vector of the interaction object stored in the recommendation model are used to determine the predicted interaction index between the target user and the interaction object, that is, the probability that the target user may interact with the interaction object, and then the target interaction object that can be recommended to the target user is determined from the plurality of interaction objects according to the probability. In this way, the target interaction object that the target user may be interested in can be recommended to the target user by only inputting the user information of the target user, which is not only simple to operate, but also extracts more interaction information between the user and the interaction object, thereby improving the accuracy of the recommendation, so that the user can browse the content of interest, and the user experience is improved.

[0087] As another example, the recommendation model can learn and store the feature vector of the interaction object during the training process, but since the recommendation model learns the feature vector of the interaction object in each data domain, and the recommendation to the target user may only be needed in a certain data domain, the identification of the data domain can also be added to the training sample during the training of the recommendation model. In this way, when the recommendation model is used to recommend the target interaction object, the identification of the data domain and the user information of the target user can be input into the recommendation model, the user information can be processed according to the domain interaction information and the cross-domain interaction information to obtain the enhanced feature vector of the target user, and then the data domain identification is used to determine the to-be-recommended interaction object belonging to the data domain, and the enhanced feature vector of the to-be-recommended interaction object and the enhanced feature vector of the target user are used to determine the predicted interaction index between the target user and the to-be-recommended interaction object, that is, the probability that the target user may interact with the to-be-recommended interaction object, and then the target interaction object that can be recommended to the target user is determined from the to-be-recommended interaction object according to the probability. In this way, the target interaction object that the target user may be interested in is determined from the to-be-recommended interaction object and recommended to the target user, which reduces the amount of calculation of the predicted interaction index and improves the recommendation efficiency, and since more interaction information between the user and the interaction object in the domain and between the domains is extracted, the accuracy of the recommendation is improved.

[0088] In the embodiments of the present specification, since the recommendation model is a machine learning model that has learned domain interaction information and cross-domain interaction information in advance, the interaction relationship between the user and the interaction object in the data domain can be considered when recommending the target interaction object to the target user, and the indirect interaction relationship between the user and the interaction object can also be considered, that is, more abundant interaction information between the user and the interaction object is considered, the target interaction object that the target user is interested in can be determined and recommended to the target user, the accuracy of object recommendation is improved, and the user experience is improved.

[0089] In some embodiments of the present specification, the recommendation model includes a graph neural network and a hypergraph neural network, and user information is input into the recommendation model. The recommendation model is used to operate the user information based on the domain interaction information and the cross-domain interaction information. The specific implementation of obtaining the target interaction object can include:

[0090] The graph neural network is used to determine the interaction feature vector of the target user according to the user information and the domain interaction information. The hypergraph neural network is used to determine the association feature vector of the target user according to the user information and the cross-domain interaction information. Based on the interaction feature vector and the association feature vector, and the interaction feature vector and the association feature vector of each interaction object in the interaction object set obtained in advance, the prediction interaction index between the target user and each interaction object is determined. The target interaction object is determined from the interaction object set based on the plurality of prediction interaction indexes.

[0091] The graph neural network is a neural network that can process graph structure data and extract low-order interaction information between nodes in the graph. For example, the graph neural network can be a GNN (Graph Neural Network), a GraphSAGE, a GCN (Graph Convolutional Network), or the like. The hypergraph neural network is a neural network that can process hypergraph structure data and extract high-order interaction information between nodes in the graph. For example, the hypergraph neural network can be a HGNN (Hypergraph Neural Networks).

[0092] In one or more embodiments of the present specification, since the graph neural network can extract low-order interaction information between nodes, the intra-domain interaction information between the user and the interaction object in the same data domain can be intuitively determined according to whether there is an edge between the user and the interaction object, and therefore, the intra-domain interaction information can be understood as low-order interaction information, and therefore, the intra-domain interaction information can be learned by the graph neural network during the training of the recommendation model. The hypergraph neural network can extract high-order interaction information between nodes, and the indirect interaction relationship between the user and the interaction object between different data domains cannot be intuitively obtained according to whether there is an edge between the user and the interaction object, but needs to be determined according to the cross-domain association between users in different data domains and the interaction relationship between the user and the interaction object in the same data domain, and therefore, the cross-domain interaction information can be understood as high-order interaction information, and therefore, the cross-domain interaction information can be learned by the hypergraph neural network during the training of the recommendation model. In this case, the intra-domain interaction information can be used as the parameter of the graph neural network, and the cross-domain interaction information can be used as the parameter of the hypergraph neural network.

[0093] In one or more embodiments of the present specification, the recommendation model also learns the interaction feature vectors and the association feature vectors of the plurality of interaction objects when the recommendation model is trained, and therefore, the interaction feature vectors and the association feature vectors of the interaction objects can be directly obtained during the process of recommending the target interaction object.

[0094] In some embodiments, the original feature vector of the target user is obtained according to the user information, and then the original feature vector of the target user is input into the graph neural network to determine the interaction feature vector of the target user according to the original feature vector and the intra-domain interaction information, and the original feature vector of the target user is input into the hypergraph neural network to determine the association feature vector of the target user according to the original feature vector and the cross-domain interaction information. Based on the interaction feature vector and the association feature vector, and the interaction feature vectors and the association feature vectors of each interaction object in the interaction object set obtained in advance, the enhanced feature vector of the target user and the enhanced feature vector of each interaction object are determined, the prediction interaction index between the target user and each interaction object, i.e., the probability that the user may interact with each interaction object, is determined according to the enhanced feature vectors of both, the interaction object with a prediction interaction index greater than a preset threshold is determined as the target interaction object, or the interaction object ranked in the top N interaction objects is determined as the target interaction object.

[0095] As an example, the training method of the recommendation model can be seen from the following Figure 5Corresponding description of the related embodiments. Since the recommendation model learns the intra-domain interaction information and the cross-domain interaction information in the process of training the recommendation model, the intra-domain interaction information can be understood as the structural relationship of the nodes in the intra-domain interaction graph, and the position of the target user in the intra-domain interaction graph can be determined according to the intra-domain interaction information. Then, according to the feature information of the target user and the feature information of the neighbor nodes thereof, the interaction feature vector of the target user can be determined; the cross-domain interaction information can be understood as the structural relationship of the nodes in the cross-domain interaction graph, and the position of the target user in the cross-domain interaction graph can be determined according to the cross-domain interaction information. Then, according to the feature information of the target user and the feature information of the reference nodes corresponding to the same hyperedge thereof, the association feature vector of the target user can be determined. Similarly, in the case of inputting the target user and the interaction object into the recommendation model in groups, the same processing is performed on the interaction object to determine the association feature vector and the interaction feature vector of the interaction object. Then, according to the interaction feature vector and the association feature vector of the target user, and the interaction feature vector and the association feature vector of the interaction object, the predicted interaction index between the target user and the interaction object is determined, and then the target interaction object is determined.

[0096] In the embodiments of the present specification, the interaction feature vector of the target user is determined according to the graph neural network, the association feature vector of the target user is determined according to the hypergraph neural network, and then the target interaction object is recommended for the target user according to the interaction feature vector and the association feature vector of the target user, and the interaction feature vector and the association feature vector of the interaction object. That is, the intra-domain interaction information and the cross-domain interaction information of the user and the interaction object are considered in the process of object recommendation, the target interaction object is determined based on more abundant interaction information, and the accuracy of object recommendation can be improved.

[0097] In another possible implementation manner of the present specification, before determining the target interaction object based on the user information, the intra-domain interaction information and the cross-domain interaction information, the method further includes:

[0098] Obtaining an interaction object set, wherein the interaction object set includes interaction object information of a plurality of interaction objects;

[0099] Correspondingly, the specific implementation of determining the target interaction object based on the user information, the intra-domain interaction information and the cross-domain interaction information can include:

[0100] Determining the target interaction object from the interaction object set based on the user information, the interaction object information of the plurality of interaction objects, the intra-domain interaction information and the cross-domain interaction information.

[0101] The interaction object information of the interaction object can include information capable of uniquely identifying an interaction object. For example, the interaction object information can be an interaction object identifier, such as an interaction object name, an interaction object title, or the like. The interaction object information can also include attribute information of the interaction object. For example, if the interaction object is an item, the interaction object information can be a price, a store, a review, or the like. If the interaction object is a song, the interaction object information can be a singer, an album, a lyricist, or the like. If the interaction object is a book, the interaction object information can be an author, a rating, or the like.

[0102] In this implementation, before recommending the target interaction object to the target user, the set of interaction objects needs to be obtained to determine the interaction object to be recommended. The enhanced feature vector of the target user and the enhanced feature vector of the interaction object are determined according to the user information, the interaction object information of the plurality of interaction objects, the pre-learned intra-domain interaction information, and the cross-domain interaction information. The predicted interaction indicator between the target user and each interaction object is determined by traversing the plurality of interaction objects. Then, the target interaction object is determined from the set of interaction objects according to the predicted interaction indicator.

[0103] In some embodiments, the user information and the interaction information of the plurality of interaction objects can be input into the recommendation model. The recommendation model processes the user information according to the intra-domain interaction information and the cross-domain interaction information, and processes the interaction object information of the plurality of interaction objects according to the intra-domain interaction information and the cross-domain interaction information. The enhanced feature vector of the target user and the enhanced feature vector of the interaction object are determined. The probability of interaction between the target user and each interaction object is determined according to the enhanced feature vectors of the two. The interaction object with a higher interaction probability is determined as the target interaction object.

[0104] As an example, first, the original feature vector of the target user is obtained according to the user information coding, and the original feature vector of the interaction object is obtained according to the interaction object information coding of the interaction object, and then the original feature vector of the target user and the original feature vector of the interaction object are input into the graph neural network, the interaction feature vector of the target user and the interaction feature vector of the interaction object are determined according to the original feature vector and the domain interaction information, the original feature vector of the target user and the original feature vector of the interaction object are input into the hypergraph neural network, the association feature vector of the target user and the association feature vector of the interaction object are determined according to the original feature vector and the cross-domain interaction information, and the enhanced feature vector of the target user and the enhanced feature vector of each interaction object in the interaction object set are determined based on the interaction feature vector and the association feature vector of the target user and the interaction feature vector and the association feature vector of each interaction object in the interaction object set, the prediction interaction index between the target user and each interaction object is determined according to the enhanced feature vectors of both, that is, the probability that the user may interact with each interaction object, and the interaction object whose prediction interaction index is greater than a preset threshold is determined as the target interaction object, or the interaction object whose prediction interaction index ranks in the top N is determined as the target interaction object.

[0105] wherein N is a positive integer greater than 0, and N can be set and adjusted according to actual conditions, which is not limited in the embodiments of the present application.

[0106] In some other embodiments, the user and the interaction object can be input into the recommendation model as an input group, so that the recommendation model can determine that the prediction interaction index to be calculated is between the user and the interaction object. In this case, any user and any interaction object can be input into the recommendation model as an input group, and the recommendation model can still correctly determine which user and which interaction object to predict the probability of interaction between. For example, user A and interaction object a are taken as a group, user A and interaction object b are taken as a group, user B and interaction object c are taken as a group, and user C and interaction object a are taken as a group, and the four groups are input into the recommendation model at the same time, then the recommendation model can output the probability of interaction between user A and interaction object a, the probability of interaction between user A and interaction object b, the probability of interaction between user B and interaction object c, and the probability of interaction between user C and interaction object a.

[0107] In the embodiments of the present application, the user information of the target user and the interaction object information of the plurality of interaction objects in the interaction object set are input into the recommendation model, and the user information and the interaction object are processed according to the pre-learned domain interaction information and cross-domain interaction information, which can enable the recommendation model to clearly determine the probability of interaction between which user and which interaction object, and further determine the target interaction object for the target user from the interaction object set, and considering the domain interaction information and the cross-domain interaction information, the accuracy of object recommendation can be improved.

[0108] Step 306: recommending the target interactive object to the target user.

[0109] In an embodiment of the present specification, the target interactive object can be sent to the client, and the client displays the target interactive object, that is, the target interactive object is recommended to the target user.

[0110] Referring to Figure 4 , Figure 4 A page schematic diagram of displaying a target interactive object according to an embodiment of the present specification is shown. In Figure 4 , the client displays interactive objects a, b, c, and d, which are target interactive objects recommended to the target user. If the target user clicks the interactive object a, an interaction between the target user and the interactive object a occurs.

[0111] The object recommendation method provided by the embodiment of the present specification obtains user information of a target user, pre-learned intra-domain interaction information and cross-domain interaction information, wherein the intra-domain interaction information represents the interaction between the user and the interactive object within the data domain, and the cross-domain interaction information represents the indirect interaction between the user and the interactive object between multiple data domains; based on the user information, the intra-domain interaction information and the cross-domain interaction information, a target interactive object is determined; and the target interactive object is recommended to the target user. That is, in the present scheme, the intra-domain interaction information of the user and the interactive object within the data domain, and the cross-domain interaction information of the user and the interactive object between different data domains are pre-learned. Since the user and the interactive object have interaction within the domain, it indicates that the user is interested in the interactive object, so the interactive object can be recommended to the user. The user and the interactive object have an indirect interaction relationship between different data domains, which indicates that the user and the interactive object may not have interaction at present, but according to the cross-domain association between users, the cross-domain association between interactive objects, and the intra-domain interaction relationship between the user and the interactive object, it can be determined that the user and the interactive object are likely to have interaction, so the interactive object can be recommended when the object is recommended to the user. The present scheme extracts higher-order and deeper cross-domain interaction information between the user and the interactive object, and considers more abundant interaction information when recommending the object, so that the target interactive object determined is more in line with the preferences of the target user, improving the accuracy of object recommendation. And in the case of sparse data in a certain data domain, since the indirect interaction relationship between the user and the interactive object between different data domains is extracted, accurate object recommendation can be realized for the user in the data domain based on the indirect interaction relationship.

[0112] In the embodiment of the present specification, the recommendation model can include a graph neural network and a hypergraph neural network. Before implementing object recommendation based on the recommendation model, the recommendation model needs to be trained, that is, before the step of inputting the user information into the recommendation model in step 304, the following recommendation model training method needs to be performed to obtain the recommendation model.

[0113] Referring to Figure 5 , Figure 5 A flowchart of a recommendation model training method according to one embodiment of the present specification is shown, which can specifically include the following steps.

[0114] Step 502: Obtain a plurality of sample groups of at least two data domains and an interaction label corresponding to each sample group.

[0115] The sample group includes a user and an interaction object, and the interaction label is used to represent whether there is an interaction relationship between the user and the interaction object, and the interaction label includes two kinds of 1 and 0, 1 represents that there is an interaction relationship, and 0 represents that there is no interaction relationship.

[0116] As an example, the plurality of sample groups can be determined according to the historical interaction data of the user in the at least two data domains, and the sample group can include a positive sample group and a negative sample group. The interaction label corresponding to the positive sample group is 1, indicating that the user in the sample group has an interaction relationship with the interaction object, and the interaction label corresponding to the negative sample group is 0, indicating that the user in the sample group has no interaction relationship with the interaction object.

[0117] In one or more embodiments of the present specification, for any data domain, the historical interaction data of a plurality of users in the data domain can be obtained, which can include operations such as browsing, following, liking, commenting, purchasing, playing, shielding, reporting, and unfollowing performed by the user on the interaction object. The operations can be divided into positive operations and negative operations, the positive operation refers to the operation indicating that the user is interested, and the negative operation refers to the operation indicating that the user is not interested. Then, the interaction object corresponding to the positive operation performed by the user and the user are grouped as a group to obtain a sample group, and the sample group is a positive sample group, and the corresponding interaction label is 1. The historical interaction data of the user can also be used to determine the interaction object that the user is not interested in, and the user and the interaction object that the user is not interested in are grouped as a group to obtain a sample group, and the sample group is a negative sample group, and the corresponding interaction label is 0.

[0118] As an example, the interaction object on which the user performs the negative operation and the user can also be grouped as a group to obtain a sample group, and the sample group is a negative sample group, and the corresponding interaction label is 0. In this way, not only can the sample group be enriched, but also the probability of recommending the interaction object that the user is not interested in to the user can be reduced.

[0119] For example, in the music playing scenario, the data domain can be the music playing field, and the interaction data of the user in the data domain can include: user A purchases and plays songs a and b, collects song c, user B plays song a, cancels the collection of song c, user C comments on song c, and cancels the collection of song e. It can be determined that user A and song a form a positive sample group, user A and song b form a positive sample group, user A and song c form a positive sample group, user B and song a form a positive sample group, user B and song b form a positive sample group, user C and song c form a positive sample group, user B and song c form a negative sample group, and user C and song e form a negative sample group. If neither user A nor user B interacts with song e, user A and song d form a negative sample group, and user B and song e form a negative sample group.

[0120] The above method can obtain the historical interaction data of the user in multiple data domains, construct sample groups based on the historical interaction data of the user, and determine the interaction labels of the sample groups, thereby providing a data basis for the training of a subsequent recommendation model.

[0121] Step 504: Based on the multiple sample groups, construct an intra-domain interaction graph corresponding to each data domain respectively, and a cross-domain interaction graph between different data domains.

[0122] In one or more embodiments of the present specification, the intra-domain interaction graph is constructed according to the interaction relationship between the user and the interaction object in the data domain, and an edge must exist for an interaction relationship to exist. Therefore, the intra-domain interaction graph is constructed according to the positive sample group. Whether an edge exists between two nodes of the same type in the cross-domain interaction graph depends on whether the two nodes interact with the same node of another type. Therefore, the cross-domain interaction graph is also constructed according to the positive sample group.

[0123] In an embodiment of the present specification, the specific implementation of constructing an intra-domain interaction graph corresponding to each data domain respectively and a cross-domain interaction graph between different data domains based on multiple sample groups can include:

[0124] According to the users and interaction objects included in the multiple sample groups in the same data domain and the interaction labels corresponding to each sample group, an intra-domain interaction graph corresponding to the data domain is constructed. If it is determined according to the multiple sample groups of at least two data domains that multiple first samples and a same second sample have an interaction relationship, a cross-domain interaction graph is constructed according to the feature information of the multiple first samples, wherein the first sample and the second sample are of different types, and the first sample is a user or an interaction object.

[0125] As an example, for the first data domain, the users and the interaction objects in the first data domain can be taken as nodes, the positive sample groups are determined according to the interaction labels of each sample group, and edges are constructed between the nodes corresponding to the users and the interaction objects having the interaction relationship in the positive sample groups, so as to construct the intra-domain interaction graph corresponding to the first data domain.

[0126] Exemplarily, it is assumed that the first data domain includes 4 positive sample groups, the first positive sample group includes user A and interaction object a1, the second positive sample group includes user B and interaction object b1, the third positive sample group includes user C and interaction object a1, and the fourth positive sample group includes user A and interaction object c1. The intra-domain interaction graph corresponding to the first data domain can be constructed by taking user A, user B, user C, interaction object a1, interaction object b1 and interaction object c1 as nodes. See Figure 6 , Figure 6 A schematic diagram of an intra-domain interaction graph provided by an embodiment of the present specification is shown, Figure 6 which is a schematic diagram of the intra-domain interaction graph corresponding to the first data domain.

[0127] In some embodiments of the present specification, if the first sample is a user, the second sample is an interaction object, that is, there are multiple users having an interaction relationship with the same interaction object, it is considered that there is an association relationship between the multiple users, and the cross-domain interaction graph is constructed from the perspective of the interaction object, representing the cross-domain association relationship between the users.

[0128] Exemplarily, see Figure 7 , Figure 7 A schematic diagram of a cross-domain interaction graph provided by an embodiment of the present specification is shown. In Figure 7 , the nodes A, B, C and D are all user nodes, user A of data domain 1, user B of data domain 2 and user C of data domain 3 all interact with interaction object a, user A of data domain 1, user C of data domain 3 and user D of data domain 4 all interact with interaction object b, there is a hyperedge e1 between nodes A, B and C, and there is a hyperedge e2 between nodes A, C and D, so that the cross-domain interaction graph representing the association relationship between the users of different data domains from the perspective of the interaction object can be constructed.

[0129] In other embodiments of the present specification, if the first sample is an interaction object, the second sample is a user, that is, there are multiple interaction objects having an interaction relationship with the same user, it is considered that there is an association relationship between the multiple interaction objects, and the cross-domain interaction graph is constructed from the perspective of the user, representing the cross-domain association relationship between the interaction objects.

[0130] Exemplarily, see Figure 8 , Figure 8Fig. 2 shows a schematic diagram of another cross-domain interaction graph according to an embodiment of the present specification. Figure 8 In Fig. 2, the interaction object a of the data domain 1, the interaction object b of the data domain 2 and the interaction object c of the data domain 3 all interact with the user A, the interaction object a of the data domain 1, the interaction object c of the data domain 3 and the interaction object d of the data domain 4 all interact with the user B, then there is a hyperedge e1 between the nodes a, b and c, and there is a hyperedge e2 between the nodes a, c and d, so as to construct a cross-domain interaction graph representing the association relationship between the interaction objects of different data domains from the perspective of the user.

[0131] As an example, the intra-domain interaction graph corresponding to each data domain can be constructed first, then the same interaction objects in different data domains are merged from the perspective of the interaction object, so as to determine the users having the association relationship, or the same users in different data domains are merged from the perspective of the user, so as to determine the interaction objects having the association relationship, and then the cross-domain interaction graph is constructed.

[0132] As an example, from the perspective of the interaction object, the same interaction objects in the intra-domain interaction graphs of two data domains can be merged according to the intra-domain interaction graphs of the two data domains, it can be determined which users interact with each interaction object, and then it can be determined which users have the association relationship, and the cross-domain interaction graph is obtained. Referring to Fig. 1, the same interaction objects in the intra-domain interaction graphs of the data domain 1 and the data domain 2 are merged, it can be determined which users interact with each interaction object, and then it can be determined which users have the association relationship, and the cross-domain interaction graph is obtained. Figure 9 , Figure 9 Fig. 3 shows a schematic diagram of still another cross-domain interaction graph according to an embodiment of the present specification. Figure 9 In Fig. 3, the nodes A, B, C, D, E, F and G are all user nodes, and the users A, B and E belong to the data domain 1, the users C, H and G belong to the data domain 2, the users D and F belong to the data domain 3, the users A, C and F have the interaction relationship with the same interaction object, then there is a hyperedge e1 between the nodes A, C and F, the users B, D and H have the interaction relationship with the same interaction object, then there is a hyperedge e2 between the nodes B, D and H, the users D, G and E have the interaction relationship with the same interaction object, then there is a hyperedge e3 between the nodes D, G and E.

[0133] In the present scheme, only the cross-domain interaction graph representing the cross-domain association relationship between the users can be constructed, or only the cross-domain interaction graph representing the cross-domain association relationship between the interaction objects can be constructed, or both the two kinds of cross-domain interaction graphs can be constructed.

[0134] In the embodiments of the present specification, the intra-domain interaction graph corresponding to each data domain and the cross-domain interaction graph between different data domains are constructed, which provides a data basis for subsequent extraction of the intra-domain interaction information and the cross-domain interaction information.

[0135] Step 506: for each data field, using a graph neural network, the feature information of the current node and the feature information of the neighbor nodes in the intra-domain interaction graph corresponding to the data field are aggregated to obtain the interaction feature vector of the current node.

[0136] wherein the current node is any node in the intra-domain interaction graph, the current node is a user or an interaction object, and the neighbor node is a node having an interaction relationship with the current node.

[0137] wherein the feature information can include the node identifier of the node, and can also include the attribute information of the node. If the node is a user, the feature information can be the user identifier or the attribute information of the user, and if the node is an interaction object, the feature information can be the interaction object identifier or the attribute information of the interaction object.

[0138] In some embodiments, the intra-domain interaction graph of each data field can be input into the graph neural network, and the initial feature vector of each node can be obtained by encoding the feature information of each node in the intra-domain interaction graph, and then the initial feature vector of the current node and the initial feature vectors of the neighbor nodes are aggregated to obtain the interaction feature vector of the current node.

[0139] For example, referring to Figure 6 with the user A as the current node, the initial feature vector of the user A is determined according to the feature information of the user A, the initial feature vector of the interaction object a is determined according to the feature information of the interaction object a, and the initial feature vector of the interaction object c is determined according to the feature information of the interaction object c, and then the initial feature vector of the user A, the initial feature vector of the interaction object a and the initial feature vector of the interaction object c are aggregated by the aggregation function to obtain the interaction feature vector of the user A, and so on. The interaction feature vector of each node in the intra-domain interaction graph can be determined.

[0140] As an example, the feature information of the nodes in the intra-domain interaction graph can be encoded to obtain the initial feature vector of the nodes, the neighbor nodes of the current node in the intra-domain interaction graph are sampled by the neighbor sampling function, the initial feature vectors of the sampled neighbor nodes are aggregated, the aggregation result is combined with the output vector of the last iteration and is subjected to convolution processing, the convolution result is subjected to nonlinear processing, and after several iterations, the iteration result is subjected to normalization processing, and the result of the normalization processing is determined as the interaction feature vector of the current node.

[0141] In the embodiments of this specification, a graph neural network is used to process the domain interaction graph to extract the domain interaction information between the user and the interaction object in the domain interaction graph, and to determine the interaction feature vector between the user and the interaction object. This interaction feature vector is obtained after learning the domain interaction information. It can be understood that this interaction feature vector can not only represent the attribute information of the user or the interaction object itself, but also represent the domain interaction information between the user and the interaction object. That is, it extracts the low-level interaction information between the user and the interaction object, providing more data references for subsequent object recommendations.

[0142] Step 508: Using a hypergraph neural network, aggregate the feature information of the reference node and the target node in the cross-domain interaction graph to obtain the associated feature vector of the reference node.

[0143] In this context, the reference node is any node in the cross-domain interaction graph, and the target node is a node that has a hyperedge with the reference node.

[0144] In some embodiments, the cross-domain interaction graph can be input into a hypergraph neural network. First, the initial feature vector of each node is obtained by encoding the feature information of each node in the cross-domain interaction graph. Then, the feature vectors of the reference node and the feature vector of the target node in the cross-domain interaction graph are aggregated to obtain the associated feature vector of the reference node.

[0145] As an example, we can first determine the hyperedge corresponding to the reference node, then determine the feature vector of the hyperedge based on the feature vectors of the other nodes corresponding to the hyperedge and the reference node, and finally aggregate the feature vectors of all the hyperedges corresponding to the reference node to obtain the associated feature vector of the reference node.

[0146] For example, see Figure 10 , Figure 10 This diagram illustrates a method for determining associated feature vectors based on cross-domain interaction graphs, according to one embodiment of this specification. Figure 10In the specific implementation, the nodes A1, B2, C3 and D4 are all user nodes, and the user A1 belongs to the data domain 1, the user B2 belongs to the data domain 2, the user C3 belongs to the data domain 3, and the user D4 belongs to the data domain 4; the user A1 of the data domain 1, the user B2 of the data domain 2 and the user C3 of the data domain 3 have an interaction relationship with the same interaction object, and the user A1 of the data domain 1, the user C3 of the data domain 3 and the user D4 of the data domain 4 have an interaction relationship with the same interaction object; the nodes A1, B2 and C3 have the hyperedge e1 therebetween, and the nodes A1, C3 and D4 have the hyperedge e2 therebetween; the feature vector of the hyperedge e1 can be determined according to the initial feature vectors of the nodes A1, B2 and C3, the feature vector of the hyperedge e2 can be determined according to the initial feature vectors of the nodes A1, C3 and D4, and the associated feature vectors of the nodes A1, B2, C3 and D4 can be determined according to the feature vector of the hyperedge e1 and the feature vector of the hyperedge e2.

[0147] For example, taking the cross-domain interaction graph as an example. Referring to FIG. 2, the cross-domain interaction graph is used to represent the cross-domain association relationship between interaction objects. Figure 7 Taking the user A as a reference node, the initial feature vector of the user A is determined according to the feature information of the user A, the initial feature vector of the user B is determined according to the feature information of the user B, the initial feature vector of the user C is determined according to the feature information of the user C, the initial feature vector of the user D is determined according to the feature information of the user D, the initial feature vector of the hyperedge e1 is determined according to the initial feature vector of the user A, the initial feature vector of the user B and the initial feature vector of the user C, the initial feature vector of the hyperedge e2 is determined according to the initial feature vector of the user A, the initial feature vector of the user C and the initial feature vector of the user D, the initial feature vector of the hyperedge e1 and the initial feature vector of the hyperedge e2 are aggregated to obtain the associated feature vector of the user A, and the associated feature vectors of all the nodes in the cross-domain interaction graph can be determined in the same manner.

[0148] In the embodiments of the present disclosure, the cross-domain interaction graph is processed by the hypergraph neural network to extract the association relationship between users and the association relationship between interaction objects in the cross-domain interaction graph, and obtain the associated feature vector of the user and the associated feature vector of the interaction object. The associated feature vector of the user is obtained after learning the high-order association relationship between users, and the associated feature vector of the interaction object is obtained after learning the high-order association relationship between interaction objects. That is, the present scheme extracts the high-order association relationship, and provides more abundant information for subsequent object recommendation.

[0149] Step 510: determining a predicted interaction index corresponding to each sample group based on the associated feature vector and the interaction feature vector.

[0150] The predicted interaction index is used to represent the probability of interaction between the user and the interaction object.

[0151] In one or more embodiments of the present specification, the method for determining the predicted interaction indicator is different according to the different constructed cross-domain interaction graphs. If only the interaction objects are taken as the perspective, the cross-domain interaction graph representing the cross-domain association relationship between users is constructed, and the first embodiment described below is used to determine the predicted interaction indicator; if only the users are taken as the perspective, the cross-domain interaction graph representing the cross-domain association relationship between interaction objects is constructed, and the second embodiment described below is used to determine the predicted interaction indicator; if both the two kinds of cross-domain interaction graphs are constructed, the third embodiment described below is used to determine the predicted interaction indicator.

[0152] In the first embodiment of the present specification, if the reference node is a user, the predicted interaction indicator corresponding to each sample group is determined based on the association feature vector, the interaction feature vector of each user in each data domain, and the interaction feature vector of each interaction object, including:

[0153] For each user in the first data domain, the interaction feature vector of the user and the association feature vector are spliced to obtain a merged feature vector, wherein the first data domain is any one of the at least two data domains; the merged feature vector of each user and the interaction feature vector of each interaction object are mapped to a data space with hierarchical structure expression capability to obtain the enhanced feature vector of each user and the enhanced feature vector of each interaction object in the first data domain; and the predicted interaction indicator corresponding to each sample group in the first data domain is determined based on the enhanced feature vector of each user and the enhanced feature vector of each interaction object in the first data domain.

[0154] The data space with hierarchical structure expression capability can be a hyperbolic space. The hierarchical information is used for the hierarchy between users or the hierarchy between interaction objects. Exemplarily, the hierarchical information of the user includes the activity of the user, and the hierarchical information of the interaction object can include the popularity (hotness) of the interaction object, the category structure (category), etc.

[0155] In some embodiments, the merged feature vector obtained by splicing the interaction feature vector of the user and the association feature vector can not only represent the intra-domain interaction information of the user and the interaction object in the same data domain, but also represent the cross-domain interaction information of the user and the interaction object between different data domains. In order to obtain more rich hierarchical information, the merged feature vector of each user and the interaction feature vector of each interaction object can be mapped to a hyperbolic space to obtain the enhanced feature vector of the user and the enhanced feature vector of the interaction object, the enhanced feature vector of the user including the hierarchical information of the user, and the enhanced feature vector of the interaction object including the hierarchical information of the interaction object.

[0156] Exemplarily, in the present scheme, the feature vectors that can be determined by default are all the feature vectors in the Euclidean space, and thus the mapping of the feature vectors from the Euclidean space to the hyperbolic space can be realized through the following formulas (1) and (2):

[0157]

[0158]

[0159] In formula (1), x is the feature vector in the Euclidean space, that is, the user merging feature vector and the interaction feature vector of the interaction object, and exp is the operation from the Euclidean space to the hyperbolic space; u is the reference vector in the hyperbolic space, which is generally the 0 vector; and c is the curvature constant.

[0160] In this implementation manner, the associated feature vector of the user is determined through the hypergraph neural network, the interaction feature vector of the interaction object is mapped to obtain the enhanced feature vector including the hierarchical information of the interaction object, the associated feature vector of the user is spliced with the interaction feature vector of the user, and then the enhanced feature vector including the hierarchical information of the user is mapped to obtain the enhanced feature vector of the user. The enhanced feature vector of the user contains high-order interaction information and hierarchical information between the user and the interaction object, so that the enhanced feature vector of the user is more comprehensive and accurate.

[0161] As an example, the enhanced feature vector of the user can include an enhanced interaction feature vector of the user and an enhanced associated feature vector of the user, and the enhanced feature vector of the interaction object includes an enhanced interaction feature vector of the interaction object. Exemplarily, referring to Figure 11 , Figure 11 A flowchart of processing of a feature vector according to an embodiment of the present specification is shown. In Figure 11 , ua and ia are the user and the interaction object in the data domain a, ub and ib are the user and the interaction object in the data domain b, and uc and ic are the user and the interaction object in the data domain c. The processing of the user and the interaction object in each data domain is the same. In the present scheme, taking the processing in the data domain a as an example, Gua represents the enhanced interaction feature vector of the user, Hua represents the enhanced associated feature vector of the user, HGua represents the enhanced feature vector of the user, and Gia represents the enhanced interaction feature vector of the interaction object.

[0162] In the second embodiment of the present specification, the reference node is the interaction object, and based on the associated feature vector, the interaction feature vector of each user in each data domain, and the interaction feature vector of each interaction object, the corresponding predicted interaction index of each sample group is determined, including:

[0163] The interaction feature vector of each interaction object in the first data domain is spliced with the association feature vector to obtain a merged feature vector, wherein the first data domain is any one of the at least two data domains; the interaction feature vector of each user and the merged feature vector of each interaction object are mapped to a data space with hierarchical structure expression capability to obtain an enhanced feature vector of each user and an enhanced feature vector of each interaction object in the first data domain; and the prediction interaction index corresponding to each sample group in the first data domain is determined based on the enhanced feature vector of each user and the enhanced feature vector of each interaction object in the first data domain.

[0164] As an example, the enhanced feature vector of the user and the enhanced feature vector of the interaction object can be determined according to the above formulas (1) and (2), except that x in the formulas is replaced by the interaction feature vector of the user and the merged feature vector of the interaction object.

[0165] In this implementation, the association feature vector of the interaction object is determined by the hypergraph neural network, the interaction feature vector of the user is mapped to obtain the enhanced feature vector including the hierarchical information of the user, and the association feature vector of the interaction object is spliced with the interaction feature vector of the interaction object to further map to obtain the enhanced feature vector including the hierarchical information of the interaction object. Therefore, the enhanced feature vector of the interaction object contains high-order interaction information and hierarchical information between the user and the interaction object, so that the enhanced feature vector of the interaction object is more comprehensive and accurate.

[0166] In the third embodiment of the present specification, not only the association feature vector of the user is determined, but also the association feature vector of the interaction object is determined. Therefore, the prediction interaction index corresponding to each sample group is determined based on the association feature vector, the interaction feature vector of each user and the interaction feature vector of each interaction object in each data domain, which includes:

[0167] The interaction feature vector of each user in the first data domain is spliced with the association feature vector to obtain a merged feature vector, wherein the first data domain is any one of the at least two data domains; the interaction feature vector of each interaction object in the first data domain is spliced with the association feature vector to obtain a merged feature vector; the merged feature vector of each user and the merged feature vector of each interaction object are mapped to a data space with hierarchical structure expression capability to obtain an enhanced feature vector of each user and an enhanced feature vector of each interaction object in the first data domain; and the prediction interaction index corresponding to each sample group in the first data domain is determined based on the enhanced feature vector of each user and the enhanced feature vector of each interaction object in the first data domain.

[0168] As an example, the enhanced feature vector of the user and the enhanced feature vector of the interactive object can be determined according to the above formulas (1) and (2), only that x in the formulas is replaced by the merged feature vector of the user and the merged feature vector of the interactive object.

[0169] In this implementation, the associated feature vector of the user and the associated feature vector of the interactive object are determined by the hypergraph neural network respectively, then the associated feature vector of the interactive object is spliced with the interaction feature vector of the interactive object, and then an enhanced feature vector including the hierarchical information of the interactive object is mapped, the interaction feature vector and the associated feature vector of the user are spliced, and then an enhanced feature vector including the hierarchical information of the user is mapped, in this way, the enhanced feature vector of the interactive object and the enhanced feature vector of the user both contain high-order interaction information and hierarchical information between the user and the interactive object, that is, both of the two enhanced feature vectors are more comprehensive and accurate.

[0170] In some embodiments, the recommendation model further includes a feature extraction layer and a fully connected layer, and the prediction interaction index corresponding to each sample group in the first data domain is determined based on the enhanced feature vector of each user and the enhanced feature vector of each interactive object in the first data domain, including:

[0171] The enhanced feature vector of each user and the interaction feature vector of the user are spliced, the spliced feature vector of each user is input into the feature extraction layer to obtain a fusion feature vector of each user, the enhanced feature vector of each interactive object is input into the feature extraction layer to obtain a fusion feature vector of each interactive object, and the fusion feature vectors of the users and the fusion feature vectors of the interactive objects belonging to the same sample group are spliced and input into the fully connected layer to determine the prediction interaction index corresponding to the sample group.

[0172] The feature extraction layer is configured to perform linear mapping, nonlinear transformation and feature extraction processing on the input feature vector. The fully connected layer is configured to perform linear mapping, nonlinear transformation and normalization processing on the input feature vector to obtain a one-dimensional feature vector, and the one-dimensional feature vector can be converted into the prediction interaction index.

[0173] As an example, the enhanced feature vector of the user can also be dimensionally adjusted to obtain a dimensionally adjusted enhanced feature vector, and then the dimensionally adjusted enhanced feature vector and the interaction feature vector of the user are spliced.

[0174] For example, referring to Figure 11, the enhanced feature vector of the user is represented as HGua, the enhanced feature vector Dua is obtained by performing dimension adjustment on HGua, the interaction feature vector Gua of the user is spliced with Dua, the spliced feature vector is input into the feature extraction layer to obtain the fusion feature vector Eua of the user, the enhanced interaction feature vector Gia of the interaction object is input into the feature extraction layer to obtain the fusion feature vector Eia of the interaction object, the fusion feature vector Eua of the user and the fusion feature vector Eia of the interaction object belonging to the same sample group are spliced and input into the full connection layer to determine the predicted interaction index ra corresponding to the sample group. The loss value is determined according to ra of a plurality of sample groups in each data field and the interaction label reala corresponding to each sample group, and the recommendation model is trained until the training stop condition is reached.

[0175] In the embodiments of the present specification, the predicted interaction index of the user and the interaction object in the sample group is determined through the processing of the feature extraction layer and the full connection layer. Since the input enhanced feature vector includes not only the domain interaction information of the user and the interaction object, but also the cross-domain interaction information of the user and the interaction object, and also includes the hierarchical information of the user and the hierarchical information of the interaction object, that is, the probability of interaction between the user and the interaction object is determined according to rich information, more accurate results can be obtained.

[0176] In some embodiments, before determining the predicted interaction index corresponding to each sample group based on the association feature vector, the interaction feature vector of each user in each data field and the interaction feature vector of each interaction object, the method further includes:

[0177] The users and the interaction objects included in each sample group are encoded to obtain the original feature vector of each user and the original feature vector of each interaction object.

[0178] The predicted interaction index corresponding to each sample group is determined based on the association feature vector, the interaction feature vector of each user in each data field and the interaction feature vector of each interaction object, including:

[0179] The predicted interaction index corresponding to each sample group is determined based on the association feature vector, the interaction feature vector and the original feature vector of each user in each data field, and the interaction feature vector and the original feature vector of each interaction object.

[0180] As an example, the users and the interaction objects included in the sample groups can be one-hot encoded first to determine the original feature vectors of the users and the original feature vectors of the interaction objects, i.e., one-hot embedding. Then, in subsequent feature vector processing, the original feature vectors are spliced in, and then the previous manner is used to determine the predicted interaction indicators. In this way, more information of the users and the interaction objects can be retained, the enhanced interaction features of the determined users and the interaction objects are more comprehensive, and the accuracy of the predicted interaction indicators corresponding to each sample group determined is improved.

[0181] In some embodiments, the original feature vectors can also be used as inputs of the feature extraction layer, so that the feature extraction layer can extract the most initial feature information of the users or the interaction objects, and then obtain the fusion feature vectors including richer and more accurate information, improve the accuracy of the determined predicted interaction indicators, and further improve the accuracy of the object recommendation.

[0182] Step 512: Based on the predicted interaction indicators corresponding to each sample group and the interaction labels corresponding to each sample group, the parameters of the recommendation model are adjusted until a training stopping condition is reached, and the recommendation model is obtained.

[0183] In some embodiments, the plurality of sample groups include positive sample groups and negative sample groups; based on the predicted interaction indicators corresponding to each sample group and the interaction labels corresponding to each sample group, the parameters of the recommendation model are adjusted until a training stopping condition is reached, and the recommendation model is obtained, including:

[0184] determining the first predicted interaction indicators corresponding to the positive sample groups and the second predicted interaction indicators corresponding to the negative sample groups in each data domain; determining the loss values corresponding to each data domain through a loss function according to the first predicted interaction indicators and the second predicted interaction indicators; and adjusting the parameters of the recommendation model according to the loss values corresponding to the at least two data domains until a training stopping condition is reached, and obtaining the recommendation model.

[0185] As an example, for the first data domain, the first predicted interaction indicators corresponding to the positive sample groups and the second predicted interaction indicators corresponding to the negative sample groups in the first data domain can be determined according to whether each sample group is a positive sample group or a negative sample group, and the loss value corresponding to the first data domain can be determined through the following formula (3) according to the first predicted interaction indicators and the second predicted interaction indicators:

[0186] loss = -sum (log (sigmoid (ri-rj))) (3)

[0187] wherein, ri represents the first predicted interaction indicator corresponding to a positive sample pair, and rj represents the second predicted interaction indicator corresponding to a negative sample pair.

[0188] In some embodiments, after determining the loss value of each data field, the plurality of loss values are summed to obtain a loss sum, if the loss sum is greater than a loss threshold, the parameters of the recommendation model are adjusted based on the loss sum, and the step of determining the interaction feature vector of the target node is returned to execute, if the loss sum is less than or equal to the loss threshold, the adjustment of the parameters of the recommendation model is stopped, and the trained recommendation model is obtained.

[0189] As an example, in one or more embodiments of the present specification, the loss can be determined and the parameters of the recommendation model can be optimized by a BPR (Bayesian Personalized Ranking) method to obtain the trained recommendation model.

[0190] The object recommendation method provided by the embodiments of the present specification first trains to obtain a recommendation model, that is, determines the association feature vector, the interaction feature vector of the user and the interaction feature vector of the interaction object according to the sample group and the interaction label corresponding to the sample group, and learns the cross-domain interaction information in the process of determining the association feature vector and learns the intra-domain interaction information in the process of determining the interaction feature vector, and then determines the predicted interaction index between the user and the interaction object in each sample group according to the association feature vector and the interaction feature vector, adjusts the parameters of the recommendation model according to the predicted interaction index and the interaction label, until the training stop condition is reached, and the training of the recommendation model is completed. The obtained user information is input into the recommendation model, and the user information is processed according to the intra-domain interaction information and the cross-domain interaction information to determine the target interaction object recommended for the target user. That is, when performing object recommendation, not only the interaction object that has actually interacted with the user is considered, but also the interaction object that may interact with the user is determined according to the cross-domain interaction information, and the target interaction object recommended based on the more comprehensive information will be more in line with the interest of the target user, that is, the accuracy of the result of the object recommendation is higher.

[0191] Referring to Figure 12 , Figure 12 A flowchart of a recommendation model training method according to an embodiment of the present specification is shown.

[0192] A plurality of sample groups of at least two data domains and corresponding interaction labels of each sample group are obtained, positive sample groups and negative sample groups are determined according to the interaction labels, a heterogeneous graph (in-domain interaction graph) of each data domain is constructed according to the interaction relationship of users and items (interaction objects) in the positive sample groups in each data domain, a user-user cross-domain hypergraph is constructed according to a plurality of users who interact with the same item between different data domains, an item-item cross-domain hypergraph is constructed according to a plurality of items that interact with the same user between different data domains, an interaction feature vector (10 dimensions) of a user is determined after learning high-order association information of the user-user cross-domain hypergraph, an interaction feature vector (10 dimensions) of an item is determined after learning high-order association information of the item-item cross-domain hypergraph, an association feature vector (10 dimensions) of a user is obtained after learning low-order interaction information of the user-item in the heterogeneous graph of each data domain, and an association feature vector (10 dimensions) of an item is obtained after learning low-order interaction information of the user-item, the interaction feature vector of the user and the association feature vector of the user are spliced to obtain a merged feature vector (20 dimensions) of the user after learning low-order interaction information and high-order interaction information of the user-item, the interaction feature vector of the interaction object and the association feature vector of the interaction object are spliced to obtain a merged feature vector (20 dimensions) of the item after learning low-order interaction information and high-order interaction information of the user-item, the merged feature vector of the user and the merged feature vector of the item are mapped to hyperbolic space to obtain an enhanced feature vector (20 dimensions) of the user after learning low-order interaction information, high-order interaction information and hierarchical information of the user-item, and an enhanced feature vector (20 dimensions) of the item after learning low-order interaction information, high-order interaction information and hierarchical information of the item, the enhanced feature vector of the user and the association feature vector of the user are spliced and input into a feature extraction layer to obtain a fusion feature vector of the user, the enhanced feature vector of the item and the association feature vector of the item are spliced and input into a feature extraction layer to obtain a fusion feature vector of the item, the fusion feature vector of the user and the fusion feature vector of the item are input into a fully connected layer to obtain a predicted interaction index, and a recommendation model is trained according to the predicted interaction index and the interaction label.

[0193] The following description is combined with the accompanying Figure 13 The object recommendation method provided in the specification is further described by taking the application of the object recommendation method in the traffic marketing scene as an example. Among them, Figure 13 A processing process flowchart of an object recommendation method applied to a traffic marketing scene is shown, which specifically includes the following steps.

[0194] Step 1302: Obtain sample groups of a plurality of data domains and corresponding interaction labels of each sample group.

[0195] The data field can include loans, financial management, traffic marketing, etc. In the data field corresponding to the traffic marketing scenario, the sample group includes users and marketing services, and the interaction label indicates whether the user handles the marketing service; in the data field corresponding to the financial management scenario, the sample group includes users and financial products, and the interaction label indicates whether the user purchases the financial product; in the data field corresponding to the loan scenario, the sample group includes users and loan services, and the interaction label indicates whether the user handles the loan service.

[0196] Step 1304: According to the sample group in the data field corresponding to the traffic marketing scenario, a first heterogeneous graph corresponding to the traffic marketing scenario is constructed, according to the sample group in the data field corresponding to the financial management scenario, a second heterogeneous graph corresponding to the financial management scenario is constructed, and according to the sample group in the data field corresponding to the loan scenario, a third heterogeneous graph corresponding to the loan scenario is constructed.

[0197] Step 1306: A user-user hypergraph between different data fields is constructed according to whether interaction with the same interaction object is generated.

[0198] Step 1308: The interaction feature vector of each user and the interaction feature vector of each marketing service in the first heterogeneous graph, the interaction feature vector of each user and the interaction feature vector of each financial product in the second heterogeneous graph, and the interaction feature vector of each user and the interaction feature vector of each loan service in the third heterogeneous graph are determined by using a graph neural network.

[0199] As an example, in the process of determining the interaction feature vector, the low-order interaction information between users and marketing services in the same data field (in-domain interaction information) is extracted, taking the first heterogeneous graph as an example.

[0200] Step 1310: The association feature vector of each user is determined by using a hypergraph neural network.

[0201] As an example, in the process of determining the association feature vector of the user, high-order association information between users and users in different data fields is extracted.

[0202] Step 1312: The association feature vector and the interaction feature vector of each user in the traffic marketing scenario are spliced to obtain a merged feature vector; the association feature vector and the interaction feature vector of each user in the financial management scenario are spliced to obtain a merged feature vector; and the association feature vector and the interaction feature vector of each user in the loan scenario are spliced to obtain a merged feature vector.

[0203] Step 1314: The merged feature vector of each user and the interaction feature vector of each interaction object are mapped to hyperbolic space to obtain an enhanced feature vector of each user and an enhanced feature vector of each interaction object.

[0204] The enhanced feature vector of each user includes hierarchical information of the user, and the enhanced feature vector of each interactive object includes hierarchical information of the interactive object.

[0205] Step 1316: The enhanced feature vector of each user and the interactive feature vector are spliced, the spliced feature vector is input into a feature extraction layer to obtain a fusion feature vector of each user, and the enhanced feature vector of each interactive object is input into the feature extraction layer to obtain a fusion feature vector of each interactive object.

[0206] Step 1318: The fusion feature vectors of the users and the fusion feature vectors of the interactive objects belonging to the same sample group are spliced and input into a full connection layer to obtain a predicted interaction index corresponding to the sample group.

[0207] Step 1320: A first predicted interaction index corresponding to a positive sample group and a second predicted interaction index corresponding to a negative sample group in each data field are determined.

[0208] Step 1322: According to the first predicted interaction index and the second predicted interaction index, a loss value corresponding to each data field is determined by a loss function.

[0209] Step 1324: The parameters of the recommendation model are adjusted according to the loss values corresponding to the plurality of data fields until a training stop condition is reached, and the recommendation model is obtained.

[0210] Step 1326: Four groups of user A and marketing service a, user A and marketing service b, user A and marketing service c, and user A and marketing service d are input into the recommendation model, and four predicted interaction indexes are output.

[0211] Step 1328: The marketing service corresponding to the maximum predicted interaction index in the four predicted interaction indexes is recommended to user A.

[0212] For example, it is assumed that user A has no historical interaction data in the marketing scene, but user A has purchased financial products a and b in the financial scene, and users B, C, D, E and F in the financial scene have also purchased financial products a and b. Users B, C, D, E and F have also handled marketing service c in the marketing scene. Through the above processing, it can be determined that the predicted interaction index corresponding to the marketing service c is the largest, and therefore the marketing service c can be recommended to user A.

[0213] It should be noted that steps 1302-1328 are the above Figure 1 and Figure 5 application in the traffic marketing scene, and the specific implementation can be referred to the related description of the above Figure 1 and Figure 5 corresponding embodiments, which will not be described here again.

[0214] The object recommendation method provided by the embodiments of the present specification first trains to obtain a recommendation model, that is, according to a sample group and an interaction label corresponding to the sample group, determines an association feature vector, an interaction feature vector of a user and an interaction feature vector of an interaction object, and learns cross-domain interaction information in the process of determining the association feature vector and learns in-domain interaction information in the process of determining the interaction feature vector, and then determines a predicted interaction index between the user and the interaction object in each sample group according to the association feature vector and the interaction feature vector, adjusts the parameters of the recommendation model according to the predicted interaction index and the interaction label, and stops until a training stop condition is reached, and the training of the recommendation model is completed. The user information obtained is input into the recommendation model, and the user information is processed according to the in-domain interaction information and the cross-domain interaction information to determine the target interaction object recommended for the target user. That is, when the object is recommended, not only the interaction object that has actually interacted with the user is considered, but also the interaction object that may interact with the user is determined according to the cross-domain interaction information. The target interaction object recommended based on the more comprehensive information will be more in line with the interest of the target user, that is, the accuracy of the result of the object recommendation is higher.

[0215] Corresponding to the method embodiments described above, the present specification also provides object recommendation device embodiments, Figure 14 The structure of an object recommendation device provided by one embodiment of the present specification is shown. As shown in the figure, Figure 14 The device includes:

[0216] The acquisition module 1402 is configured to acquire user information of a target user, in-domain interaction information and cross-domain interaction information learned in advance, wherein the in-domain interaction information represents the interaction between the user and the interaction object within the data domain, and the cross-domain interaction information represents the indirect interaction between the user and the interaction object between multiple data domains;

[0217] The determination module 1404 is configured to determine the target interaction object based on the user information, the in-domain interaction information and the cross-domain interaction information;

[0218] The recommendation module 1406 is configured to recommend the target interaction object to the target user.

[0219] In one or more embodiments of the present specification, the determination module 1404 is further configured to:

[0220] Obtain an interaction object set, wherein the interaction object set includes interaction object information of multiple interaction objects;

[0221] Determine the target interaction object based on the user information, the in-domain interaction information and the cross-domain interaction information, including:

[0222] Determine the target interaction object from the interaction object set based on the user information, the interaction object information of the multiple interaction objects, the in-domain interaction information and the cross-domain interaction information.

[0223] In one or more embodiments of the present specification, the intra-domain interaction information and the cross-domain interaction information are obtained by pre-training a recommendation model, and the recommendation model is a machine learning model.

[0224] The determining module 1404 is further configured to:

[0225] The user information is input into the recommendation model, and the recommendation model is used to operate the user information by using the intra-domain interaction information and the cross-domain interaction information, to obtain the target interaction object.

[0226] In one or more embodiments of the present specification, the recommendation model includes a graph neural network and a hypergraph neural network.

[0227] The apparatus further includes a model training module configured to:

[0228] Obtain a plurality of sample groups of at least two data domains and interaction labels corresponding to each sample group, wherein each sample group includes a user and an interaction object, and the interaction label is used to represent whether the user and the interaction object have an interaction relationship.

[0229] Based on the plurality of sample groups, an intra-domain interaction graph corresponding to each data domain and a cross-domain interaction graph between different data domains are constructed.

[0230] For each data domain, the graph neural network is used to aggregate the feature information of a current node and the feature information of neighbor nodes in the intra-domain interaction graph corresponding to the data domain, to obtain an interaction feature vector of the current node, wherein the current node is any node in the intra-domain interaction graph, the current node is a user or an interaction object, and the neighbor node is a node having an interaction relationship with the current node.

[0231] The hypergraph neural network is used to aggregate the feature information of a reference node and the feature information of a target node in the cross-domain interaction graph, to obtain an association feature vector of the reference node, wherein the reference node is any node in the cross-domain interaction graph, and the target node is a node having a hyperedge with the reference node.

[0232] Based on the association feature vector and the interaction feature vector, a predicted interaction indicator corresponding to each sample group is determined, wherein the predicted interaction indicator is used to represent a probability of interaction between the user and the interaction object.

[0233] Based on the predicted interaction indicator corresponding to each sample group and the interaction label corresponding to each sample group, the parameters of the recommendation model are adjusted until a training stop condition is reached, to obtain the recommendation model.

[0234] In one or more embodiments of the present specification, the model training module is further configured to:

[0235] construct an intra-domain interaction graph corresponding to the data domain according to the users and the interaction objects included in the plurality of sample groups in the same data domain and the interaction labels corresponding to each sample group;

[0236] If it is determined that a plurality of first samples have an interaction relationship with a same second sample according to a plurality of sample groups of at least two data domains, a cross-domain interaction graph is constructed according to feature information of the plurality of first samples, wherein the first sample and the second sample are of different types, and the first sample is a user or an interaction object.

[0237] In one or more embodiments of the present specification, the reference node is a user, and the model training module is further configured to:

[0238] For each user in the first data domain, the interaction feature vector and the association feature vector of the user are spliced to obtain a merged feature vector, wherein the first data domain is any one of the at least two data domains;

[0239] The merged feature vector of each user and the interaction feature vector of each interaction object are mapped to a data space with hierarchical structure expression capability to obtain an enhanced feature vector of each user and an enhanced feature vector of each interaction object in the first data domain;

[0240] Based on the enhanced feature vector of each user and the enhanced feature vector of each interaction object in the first data domain, a predicted interaction indicator corresponding to each sample group in the first data domain is determined.

[0241] In one or more embodiments of the present specification, the reference node is an interaction object, and the model training module is further configured to:

[0242] For each interaction object in the first data domain, the interaction feature vector and the association feature vector of the interaction object are spliced to obtain a merged feature vector, wherein the first data domain is any one of the at least two data domains;

[0243] The interaction feature vector of each user and the merged feature vector of each interaction object are mapped to a data space with hierarchical structure expression capability to obtain an enhanced feature vector of each user and an enhanced feature vector of each interaction object in the first data domain;

[0244] Based on the enhanced feature vector of each user and the enhanced feature vector of each interaction object in the first data domain, a predicted interaction indicator corresponding to each sample group in the first data domain is determined.

[0245] In one or more embodiments of the present specification, the recommendation model further includes a feature extraction layer and a fully connected layer, and the model training module is further configured to:

[0246] The enhanced feature vector of each user and the interaction feature vector of the user are spliced, and the spliced feature vector of each user is input into a feature extraction layer to obtain a fusion feature vector of each user;

[0247] The enhanced feature vector of each interaction object is input into the feature extraction layer to obtain a fusion feature vector of each interaction object.

[0248] The fusion feature vectors of the users and the fusion feature vectors of the interaction objects belonging to the same sample group are spliced and input into a full connection layer to determine a predicted interaction indicator corresponding to the sample group.

[0249] In one or more embodiments of the present specification, the plurality of sample groups includes a positive sample group and a negative sample group.

[0250] The model training module is further configured to:

[0251] determine a first predicted interaction indicator corresponding to the positive sample group and a second predicted interaction indicator corresponding to the negative sample group in each data domain;

[0252] determine a loss value corresponding to each data domain by a loss function according to the first predicted interaction indicator and the second predicted interaction indicator;

[0253] adjust parameters of the recommendation model according to the loss values corresponding to the at least two data domains until a training stop condition is reached to obtain the recommendation model.

[0254] In one or more embodiments of the present specification, the model training module is further configured to:

[0255] encode the users and the interaction objects included in each sample group to obtain an original feature vector of each user and an original feature vector of each interaction object;

[0256] determine a predicted interaction indicator corresponding to each sample group based on the association feature vector, the interaction feature vector of each user and the interaction feature vector of each interaction object in each data domain, including:

[0257] determine a predicted interaction indicator corresponding to each sample group based on the association feature vector, the interaction feature vector and the original feature vector of each user in each data domain, and the interaction feature vector and the original feature vector of each interaction object.

[0258] In one or more embodiments of the present specification, the recommendation model includes a graph neural network and a hypergraph neural network, and the model training module is further configured to:

[0259] determine the interaction feature vector of the target user according to the user information and the in-domain interaction information by using the graph neural network;

[0260] The supergraph neural network is used to determine a correlation feature vector of the target user according to the user information and the cross-domain interaction information.

[0261] Based on the interaction feature vector and the correlation feature vector, and the interaction feature vector and the correlation feature vector of each interaction object in the interaction object set that are acquired in advance, a predicted interaction index between the target user and each interaction object is determined.

[0262] The target interaction object is determined from the interaction object set based on the plurality of predicted interaction indexes.

[0263] The object recommendation device provided in the embodiments of the present specification obtains user information of a target user, domain-in interaction information and cross-domain interaction information learned in advance, wherein the domain-in interaction information represents interaction between an internal user and an interaction object in a data domain, and the cross-domain interaction information represents indirect interaction between a user and an interaction object in different data domains; based on the user information, the domain-in interaction information and the cross-domain interaction information, a target interaction object is determined; and the target interaction object is recommended to the target user. That is, in the present scheme, the domain-in interaction information between an internal user and an interaction object in a data domain, and the cross-domain interaction information between a user and an interaction object in different data domains are learned in advance. Since the user and the interaction object have interaction in the domain, it indicates that the user is interested in the interaction object, and the user and the interaction object may not have interaction at present, but according to the cross-domain correlation between users, the cross-domain correlation between interaction objects and the domain-in interaction relationship between the user and the interaction object, it can be determined that the user and the interaction object are likely to have interaction. Therefore, the interaction object can be recommended to the user when the object is recommended. The present scheme extracts higher-order and deeper cross-domain interaction information between the user and the interaction object, and considers more abundant interaction information when the object is recommended. Therefore, the target interaction object determined is more in line with the preference of the target user, improves the accuracy of object recommendation, and in the case of sparse data in a certain data domain, since the indirect interaction relationship between the user and the interaction object in different data domains is extracted, the indirect interaction relationship can be used to accurately recommend the object to the user in the data domain.

[0264] The above is a schematic scheme of the object recommendation device of the present embodiment. It should be noted that the technical scheme of the object recommendation device belongs to the same concept as the technical scheme of the object recommendation method described above. The technical scheme of the object recommendation device is not described in detail, and the description of the technical scheme of the object recommendation method can be referred to.

[0265] Figure 15A structural block diagram of a computing device 1500 is shown, which is provided according to one embodiment of the present specification. The components of the computing device 1500 include, but are not limited to, a memory 1510 and a processor 1520. The processor 1520 is connected with the memory 1510 through a bus 1530, and a database 1550 is used to save data.

[0266] The computing device 1500 also includes an access device 1540, which enables the computing device 1500 to communicate via one or more networks 1560. Examples of these networks include the public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 1540 can include one or more of any type of network interface (e.g., a network interface card (NIC)), wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and the like.

[0267] In one embodiment of the present specification, the above-mentioned components of the computing device 1500 and other components not shown in the above-mentioned components can be connected with each other, for example, through a bus. It should be understood that Figure 15 the above-mentioned components of the computing device 1500 and other components not shown in the above-mentioned components can be connected with each other, for example, through a bus. It should be understood that Figure 15 The structural block diagram of the computing device shown is only for the purpose of example, and is not a limitation on the scope of the present specification. Other components can be added or replaced as needed by those skilled in the art.

[0268] The computing device 1500 can be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smartphone), a wearable computing device (e.g., a smartwatch, smart glasses, etc.), or other type of mobile device, or a stationary computing device such as a desktop computer or a PC. The computing device 1500 can also be a mobile or stationary server.

[0269] The processor 1520 is configured to execute the following computer-executable instructions, which implement the steps of the object recommendation method described above.

[0270] The above is a schematic solution of the computing device of the embodiment. It should be noted that the technical solution of the computing device and the technical solution of the object recommendation method described above belong to the same concept, and the details of the technical solution of the computing device that are not described in detail can be referred to the description of the technical solution of the object recommendation method.

[0271] An embodiment of the present specification also provides a computer-readable storage medium storing computer-executable instructions, which implement the steps of the object recommendation method described above when executed by a processor.

[0272] The above is a schematic solution of the computer-readable storage medium of the embodiment. It should be noted that the technical solution of the storage medium and the technical solution of the object recommendation method described above belong to the same concept, and the details of the technical solution of the storage medium that are not described in detail can be referred to the description of the technical solution of the object recommendation method.

[0273] An embodiment of the present specification also provides a computer program, which causes a computer to perform the steps of the object recommendation method described above when the computer program is executed in the computer.

[0274] The above is a schematic solution of the computer program of the embodiment. It should be noted that the technical solution of the computer program and the technical solution of the object recommendation method described above belong to the same concept, and the details of the technical solution of the computer program that are not described in detail can be referred to the description of the technical solution of the object recommendation method.

[0275] The above describes specific embodiments of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the acts or steps recited in the claims can be performed in a different order than the order described in the embodiments and still achieve desirable results. In addition, the processes depicted in the figures do not necessarily require the particular order shown or sequential order in order to achieve the desired results. In some implementations, multitasking and parallel processing can be advantageous.

[0276] The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, software distribution medium, etc.

[0277] It should be noted that, for the foregoing method embodiments, in order to facilitate description, they are all expressed as a combination of a series of actions, but those skilled in the art should know that the embodiments of the present specification are not limited by the order of the described actions, because according to the embodiments of the present specification, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments of the present specification.

[0278] In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0279] The preferred embodiments of the present specification disclosed above are only used to help explain the present specification. The alternative embodiments do not describe all the details and limit the invention to the specific embodiments described. Obviously, according to the content of the embodiments of the present specification, many modifications and changes can be made. The present specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of the present specification, so that those skilled in the art can well understand and use the present specification. The present specification is limited by the claims and their entire scope and equivalents.

Claims

1. A method for object recommendation, comprising: obtaining user information of a target user, pre-learned intra-domain interaction information and cross-domain interaction information, wherein the intra-domain interaction information represents interactions between the target user and interaction objects within a data domain, and the cross-domain interaction information represents indirect interactions between the target user and interaction objects across multiple data domains; determining a target interaction object based on the user information, the intra-domain interaction information and the cross-domain interaction information, including: inputting the user information into a recommendation model, wherein the recommendation model comprises a graph neural network and a hypergraph neural network; determining an interaction feature vector of the target user according to the user information and the intra-domain interaction information by using the graph neural network; determining a correlation feature vector of the target user according to the user information and the cross-domain interaction information by using the hypergraph neural network; determining a predicted interaction indicator between the target user and each interaction object in an interaction object set based on the interaction feature vector and the correlation feature vector, and pre-obtained interaction feature vectors and correlation feature vectors of each interaction object in the interaction object set; and determining the target interaction object from the interaction object set based on multiple predicted interaction indicators; recommending the target interaction object to the target user.

2. The method of claim 1, before determining the target interaction object based on the user information, the intra-domain interaction information and the cross-domain interaction information, further comprising: obtaining an interaction object set, wherein the interaction object set comprises interaction object information of multiple interaction objects; determining the target interaction object based on the user information, the intra-domain interaction information and the cross-domain interaction information, including: determining the target interaction object from the interaction object set based on the user information, interaction object information of the multiple interaction objects, the intra-domain interaction information and the cross-domain interaction information.

3. The method of claim 1, before inputting the user information into the recommendation model, further comprising: obtaining multiple sample groups of at least two data domains and corresponding interaction labels of each sample group, wherein each sample group comprises a user and an interaction object, and the interaction label is used to represent whether there is an interaction relationship between the user and the interaction object; constructing an intra-domain interaction graph corresponding to each data domain and a cross-domain interaction graph between different data domains based on the multiple sample groups; for each data domain, aggregating feature information of a current node and feature information of neighbor nodes in the intra-domain interaction graph corresponding to the data domain by using the graph neural network to obtain an interaction feature vector of the current node, wherein the current node is any node in the intra-domain interaction graph, the current node is a user or an interaction object, and the neighbor nodes are nodes having an interaction relationship with the current node; aggregating feature information of a reference node and feature information of a target node in the cross-domain interaction graph by using the hypergraph neural network to obtain a correlation feature vector of the reference node, wherein the reference node is any node in the cross-domain interaction graph, and the target node is a node having a hyperedge with the reference node. ​ determine, based on the association feature vector and the interaction feature vector, a predicted interaction indicator corresponding to each sample group, where the predicted interaction indicator is used to represent a probability of interaction between a user and an interaction object; adjust parameters of the recommendation model based on the predicted interaction indicator corresponding to each sample group and the interaction label corresponding to each sample group until a training stop condition is reached, to obtain the recommendation model.

4. The method of claim 3, based on the plurality of sample groups, constructing an intra-domain interaction graph corresponding to each data domain respectively, and a cross-domain interaction graph between different data domains, comprising: constructing an intra-domain interaction graph corresponding to a data domain according to users and interaction objects included in a plurality of sample groups in the same data domain and the interaction label corresponding to each sample group; if it is determined according to the plurality of sample groups of the at least two data domains that a plurality of first samples have an interaction relationship with a same second sample, constructing a cross-domain interaction graph according to feature information of the plurality of first samples, where the first sample and the second sample are of different types, and the first sample is a user or an interaction object.

5. The method of claim 3 or 4, the reference node being a user, determining, based on the association feature vector, the interaction feature vector of each user in each data domain and the interaction feature vector of each interaction object, a predicted interaction indicator corresponding to each sample group, comprising: for each user in a first data domain, concatenating the interaction feature vector of the user and the association feature vector to obtain a merged feature vector, where the first data domain is any one of the at least two data domains; mapping the merged feature vector of each user and the interaction feature vector of each interaction object to a data space with hierarchical structure expression capability to obtain an enhanced feature vector of each user and an enhanced feature vector of each interaction object in the first data domain; determining, based on the enhanced feature vector of each user and the enhanced feature vector of each interaction object in the first data domain, a predicted interaction indicator corresponding to each sample group in the first data domain.

6. The method of claim 3 or 4, the reference node being an interaction object, determining, based on the association feature vector, the interaction feature vector of each user in each data domain and the interaction feature vector of each interaction object, a predicted interaction indicator corresponding to each sample group, comprising: for each interaction object in a first data domain, concatenating the interaction feature vector of the interaction object and the association feature vector to obtain a merged feature vector, where the first data domain is any one of the at least two data domains; mapping the interaction feature vector of each user and the merged feature vector of each interaction object to a data space with hierarchical structure expression capability to obtain an enhanced feature vector of each user and an enhanced feature vector of each interaction object in the first data domain; determining, based on the enhanced feature vector of each user and the enhanced feature vector of each interaction object in the first data domain, a predicted interaction indicator corresponding to each sample group in the first data domain.

7. The method of claim 5, wherein the recommendation model further comprises a feature extraction layer and a fully connected layer, and wherein determining the predicted interaction indicator corresponding to each sample group in the first data domain based on the enhanced feature vector of each user and the enhanced feature vector of each interaction object comprises: concatenating the enhanced feature vector of each user with the interaction feature vector of the user, and inputting the concatenated feature vector of each user into the feature extraction layer to obtain a fusion feature vector of each user; inputting the enhanced feature vector of each interaction object into the feature extraction layer to obtain a fusion feature vector of each interaction object; and inputting the fusion feature vectors of the users and the fusion feature vectors of the interaction objects belonging to the same sample group into the fully connected layer to determine the predicted interaction indicator corresponding to the sample group.

8. The method of any one of claims 3, 4, or 7, wherein the plurality of sample groups comprises positive sample groups and negative sample groups; and wherein adjusting the parameters of the recommendation model based on the predicted interaction indicator corresponding to each sample group and the interaction label corresponding to each sample group until a training stopping condition is reached to obtain the recommendation model comprises: determining a first predicted interaction indicator corresponding to the positive sample groups and a second predicted interaction indicator corresponding to the negative sample groups in each data domain; determining a loss value corresponding to each data domain based on the first predicted interaction indicator and the second predicted interaction indicator via a loss function; and adjusting the parameters of the recommendation model based on the loss values corresponding to the at least two data domains until the training stopping condition is reached to obtain the recommendation model.

9. The method of claim 3 or 4, wherein before determining the predicted interaction indicator corresponding to each sample group based on the association feature vector, the interaction feature vector of each user in each data domain, and the interaction feature vector of each interaction object, the method further comprises: encoding the users and the interaction objects included in each sample group to obtain an original feature vector of each user and an original feature vector of each interaction object; and wherein determining the predicted interaction indicator corresponding to each sample group based on the association feature vector, the interaction feature vector of each user in each data domain, and the interaction feature vector of each interaction object comprises: determining the predicted interaction indicator corresponding to each sample group based on the association feature vector, the interaction feature vector and the original feature vector of each user in each data domain, and the interaction feature vector and the original feature vector of each interaction object.

10. An object recommendation apparatus, comprising: an obtaining module configured to obtain user information of a target user, pre-learned intra-domain interaction information, and cross-domain interaction information, wherein the intra-domain interaction information represents interactions between users and interaction objects within a data domain, and the cross-domain interaction information represents indirect interactions between users and interaction objects across multiple data domains; and a determining module configured to input the user information into a recommendation model, wherein the recommendation model comprises a graph neural network and a hypergraph neural network, determine an interaction feature vector of the target user based on the user information and the intra-domain interaction information using the graph neural network, and determine an association feature vector of the target user based on the user information and the cross-domain interaction information using the hypergraph neural network. ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ The determining module is configured to determine a target interaction object based on the user information, the intra-domain interaction information, and the cross-domain interaction information ​ ​ ​ determining a correlation feature vector of the target user; determining a predicted interaction indicator between the target user and each interaction object based on the interaction feature vector and the correlation feature vector of each interaction object in a pre-acquired interaction object set and the interaction feature vector and the correlation feature vector of each interaction object in the pre-acquired interaction object set; determining a target interaction object from the interaction object set based on the predicted interaction indicators; and recommending the target interaction object to the target user.

11. A computing device, comprising: a memory and a processor; the memory is configured to store computer-executable instructions, and the processor is configured to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the object recommendation method according to any one of claims 1 to 9.

12. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the object recommendation method according to any one of claims 1 to 9. including computer instructions, which, when executed by a processor, implement the steps of the object recommendation method according to any one of claims 1 to 9. ; ​ ​ ​ ​ ​ 13. A computer program product, characterised in that, ​

Images