A multi-objective cross-domain recommendation method, device and system based on heterogeneous graphs

Abstract

The present invention is applicable to the technical field of cross-domain recommendation, and provides a multi-objective cross-domain recommendation method, device and system based on a heterogeneous graph. The multi-objective cross-domain recommendation method based on a heterogeneous graph includes: constructing an intra-domain heterogeneous graph and an inter-domain heterogeneous graph based on user information and / or item information within a domain; performing multi-layer graph convolution on all intra-domain heterogeneous graphs and inter-domain heterogeneous graphs respectively to obtain high-order representation information of each user node and / or item node; fusing the high-order representation information of the user nodes and / or item nodes to obtain fused representation information of the user nodes and / or item nodes in each domain; respectively inputting the fused representation information of all user nodes and item nodes into their respective fully connected neural networks to learn the non-linear relationship between users and items, obtaining a predicted score value of a user for an item, and performing cross-domain recommendation based on the predicted score value. This application can significantly alleviate the negative transfer problem in multi-domain recommendation and improve the recommendation accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of cross - domain recommendation, and particularly relates to a multi - target cross - domain recommendation method, device and system based on a heterogeneous graph. Background Art

[0002] In today's Internet era, the amount of information has grown geometrically, and information overload has become an issue that cannot be ignored. Recommendation systems are an effective way to solve the problem of information overload. They can recommend information and products that users are interested in to users through user and item information, enabling people to obtain the required information more efficiently. Cross - domain recommendation algorithms are one of the important branches in recommendation systems and are effective methods for solving data sparsity and cold - start problems. It aims to utilize the interaction information of users and items in multiple data domains to alleviate the data sparsity problem, thereby improving the recommendation accuracy.

[0003] In traditional recommendation systems, recommendations are usually made based on user behavior data. For example, recommendations are made based on users' historical browsing or rating information. Existing recommendation technologies are mostly limited to single - target cross - domain or dual - target cross - domain scenarios. For multi - target cross - domain recommendations, single - target cross - domain or dual - target cross - domain methods are often directly migrated to multi - target cross - domain. Such methods are difficult to accurately capture the high - order relationships between users and items, the migration relationships to be learned will increase significantly, and knowledge negative transfer is extremely likely to occur between domains, significantly affecting the accuracy of cross - domain recommendations.

[0004] Currently, there are problems of knowledge negative transfer between domains in multi - target cross - domain recommendation algorithms, and the recommendation accuracy needs to be improved. Summary of the Invention

[0005] The purpose of the embodiments of the present application is to provide a multi - target cross - domain recommendation method based on a heterogeneous graph, aiming to solve the problems that the recommendation accuracy of current multi - target cross - domain recommendation algorithms is relatively low and there is knowledge negative transfer between domains.

[0006] The embodiments of the present application are implemented as follows. A multi - target cross - domain recommendation method based on a heterogeneous graph is provided. The method includes: constructing an intra - domain heterogeneous graph and an inter - domain heterogeneous graph based on user information and / or item information within a domain; performing multi - layer graph convolution on all the intra - domain heterogeneous graphs and inter - domain heterogeneous graphs respectively to obtain high - order representation information of each user node and / or item node; fusing the high - order representation information of the user nodes and / or item nodes to obtain fused representation information of user nodes and / or item nodes in each domain; respectively inputting the fused representation information of all the user nodes and item nodes into their respective fully - connected neural networks to learn the non - linear relationship between users and items, obtaining a predicted rating value of users for items, and performing cross - domain recommendation based on the predicted rating value.

[0007] Another object of the embodiments of the present application is to provide a multi-object cross-domain recommendation device based on a heterogeneous graph. The device includes: an intra-domain heterogeneous graph and an inter-domain heterogeneous graph construction module, configured to construct an intra-domain heterogeneous graph and an inter-domain heterogeneous graph based on user information and / or item information within a domain; a high-order representation information acquisition module, configured to perform multi-layer graph convolution on all the intra-domain heterogeneous graphs and inter-domain heterogeneous graphs respectively to obtain high-order representation information of each user node and / or item node; a fused representation information acquisition module, configured to fuse the high-order representation information of the user nodes and / or item nodes to obtain fused representation information of user nodes and / or item nodes in each domain; a cross-domain recommendation module, configured to respectively input the fused representation information of all the user nodes and item nodes into their respective fully connected neural networks to learn the non-linear relationship between users and items, obtain a predicted score value of a user for an item, and perform cross-domain recommendation based on the predicted score value.

[0008] Another object of the embodiments of the present application is to provide a multi-object cross-domain recommendation system based on a heterogeneous graph. When the system runs, it executes the steps of the above-mentioned multi-object cross-domain recommendation method based on a heterogeneous graph.

[0009] A multi-object cross-domain recommendation method based on a heterogeneous graph provided by the embodiments of the present application constructs two types of heterogeneous graphs, intra-domain and inter-domain, for the problem of multi-object cross-domain recommendation, and fully learns the high-order relationship between users and items through a graph neural network to obtain accurate user and item representations; when performing node aggregation in the heterogeneous graph, a hierarchical attention mechanism is adopted to filter important information for each node, thereby alleviating the knowledge negative transfer problem in multi-object cross-domain recommendation and improving the recommendation accuracy. Description of the Drawings

[0010] Figure 1 It is an application environment diagram of a multi-object cross-domain recommendation method based on a heterogeneous graph provided by the embodiments of the present application;

[0011] Figure 2 It is a flowchart of a multi-object cross-domain recommendation method based on a heterogeneous graph provided by the embodiments of the present application;

[0012] Figure 3 It is a running flowchart of a multi-object cross-domain recommendation method based on a heterogeneous graph provided by the embodiments of the present application;

[0013] Figure 4 It is a structural block diagram of a multi-object cross-domain recommendation device based on a heterogeneous graph provided by the embodiments of the present application. Detailed Embodiments

[0014] To make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0015] It can be understood that the terms "first", "second", etc. used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are only used to distinguish one element from another. For example, without departing from the scope of this application, the first xx script may be referred to as the second xx script, and similarly, the second xx script may be referred to as the first xx script.

[0016] Figure 1 FIG. is an application environment diagram of a multi-objective cross-domain recommendation method based on a heterogeneous graph provided by an embodiment of this application. As Figure 1 shown, in this application environment, it includes a terminal 110 and a computer device 120.

[0017] The computer device 120 may be an independent physical server or terminal, or may be a server cluster composed of multiple physical servers, and may be a cloud server providing basic cloud computing services such as cloud servers, cloud databases, cloud storage, and CDN.

[0018] The terminal 110 may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto. The terminal 110 and the computer device 120 may be connected through a network, and this application does not make any restrictions here.

[0019] As an embodiment of this application, the terminal 110 is a mobile phone and the computer device 120 is a cloud server. When a user browses item information using a mobile phone, the terminal device can obtain user information and data such as item information interacted by the user through the terminal device, and send the data to the cloud server. The above-mentioned multi-objective cross-domain recommendation method based on a heterogeneous graph is executed in the cloud server, and the recommendation result is pushed to the user's mobile phone to achieve cross-domain recommendation.

[0020] As Figure 2 shown, in one embodiment, a multi-objective cross-domain recommendation method based on a heterogeneous graph is proposed. In this embodiment, this method is mainly applied to the above-mentioned Figure 1 computer device 120 in the example. A multi-objective cross-domain recommendation method based on a heterogeneous graph may specifically include the following steps:

[0021] Step S202, construct an intra-domain heterogeneous graph and an inter-domain heterogeneous graph based on the user information and / or item information within the domain.

[0022] Step S204: Perform multi-layer graph convolution on all the intra-domain heterogeneous graphs and inter-domain heterogeneous graphs respectively to obtain the high-order representation information of each user node and / or item node.

[0023] Step S206: Fuse the high-order representation information of the user nodes and / or item nodes to obtain the fused representation information of the user nodes and / or the fused representation information of the item nodes in each domain.

[0024] Step S208: Input the fused representation information of all the user nodes and the fused representation information of the item nodes into their respective fully connected neural networks to learn the non-linear relationship between users and items, obtain the predicted rating value of the user for the item, and perform cross-domain recommendation based on the predicted rating value.

[0025] In the embodiment of the present application, the multi-objective cross-domain recommendation method may include the following steps: (1) Heterogeneous graph construction: Construct intra-domain and inter-domain heterogeneous graphs according to the interaction information between users and items in each domain; (2) Graph convolution: Perform graph convolution on the intra-domain heterogeneous graphs constructed based on each domain and the inter-domain heterogeneous graphs of multiple domains, and use the hierarchical attention mechanism to aggregate messages of the nodes in the intra-domain and inter-domain heterogeneous graphs to learn the high-order representations of users and items; (3) Representation fusion: Fuse the representations of intra-domain and inter-domain users and items learned by the graph convolution layer; (4) Rating prediction: Perform non-linear feature learning on the representations of users and items to obtain the predicted rating value of the user for the item, and select recommended content based on the rating value.

[0026] In the embodiment of the present application, a method is proposed to perform graph convolution by constructing heterogeneous graphs intra-domain and inter-domain respectively, learn the high-order representations of intra-domain and inter-domain users and items, and combine the hierarchical attention mechanism to obtain the predicted rating value and perform recommendation, thereby greatly alleviating the negative transfer problem of knowledge in multi-domain recommendation and improving the cross-domain recommendation accuracy for multiple objectives.

[0027] As known to those skilled in the art, a domain may simultaneously contain several user information and item information, or one of the above information. The number of domains in the above recommendation method may be two, three or more. When the number of domains is one, the model degrades to a single-domain recommendation model. As an embodiment of the present application, here the number of domains is taken as three for example. As Figure 3 shown, the domains are a, b, and c. According to the interaction information between users and items in the three domains, intra-domain heterogeneous graphs of the three domains are constructed respectively. In the intra-domain heterogeneous graphs, there are nodes of intra-domain users and items, and each node has its own representation. According to the interaction information between users and items in the three domains, an inter-domain heterogeneous graph is constructed. In the inter-domain heterogeneous graph, there are nodes of users and items in the three domains, and each node also has its own representation. After constructing four heterogeneous graphs for the three domains, initialize the representation of each node.

[0028] In one embodiment, the method for constructing an intra-domain heterogeneous graph and an inter-domain heterogeneous graph based on user information and / or item information within a domain is as follows: construct a corresponding intra-domain heterogeneous graph for each domain, where each intra-domain heterogeneous graph contains at least one user node and / or one item node, the user information corresponds to the user node, and the item information corresponds to the item node; construct an inter-domain heterogeneous graph according to the interaction information between the user node and the item node; and there is one inter-domain heterogeneous graph.

[0029] As an embodiment of the present application, as Figure 3 shown, taking the number of domains as 3 as an example. Domains a, b, and c respectively have user and item sets Interaction matrix Construct an intra-domain heterogeneous graph according to it:

[0030]

[0031]

[0032]

[0033] Among them: respectively represent the intra-domain heterogeneous graphs of domains a, b, and c, and E is the interaction information between users and items:

[0034]

[0035]

[0036]

[0037] Construct an inter-domain heterogeneous graph:

[0038]

[0039] Among them: G inter represents the inter-domain heterogeneous graph,

[0040] In this embodiment, after constructing the intra-domain heterogeneous graphs of domains a, b, and c, each node has a k-dimensional representation vector, and Xavier initialization is used to initialize the node representation. In the inter-domain heterogeneous graph, the nodes include users and items of all domains, each node has a k-dimensional representation vector, and Xavier initialization is used to initialize the node representation. For the problem of multi-objective cross-domain recommendation, two types of heterogeneous graphs, intra-domain and inter-domain, are constructed, so as to facilitate the subsequent graph neural network to fully learn the high-order relationship between users and items and obtain accurate user and item representations.

[0041] In one embodiment, the method for performing multi-layer graph convolution on the intra-domain heterogeneous graph to obtain the high-order representation information of each user node and / or item node is as follows: perform three-layer graph convolution on each intra-domain heterogeneous graph, and use a hierarchical attention mechanism for each user node and / or item node to fuse the representation information of its neighbor nodes for message aggregation, so as to obtain the high-order representation information of all user nodes and / or item nodes in each intra-domain heterogeneous graph. The method for performing multi-layer graph convolution on the inter-domain heterogeneous graph to obtain the high-order representation information of each user node and / or item node is as follows: perform three-layer graph convolution on the inter-domain heterogeneous graph, and use a hierarchical attention mechanism for each user node and / or item node to perform message aggregation, so as to obtain the high-order representation information of all user nodes and / or item nodes in each inter-domain heterogeneous graph.

[0042] In the embodiment of the present application, according to the heterogeneous graph constructed in the foregoing steps, three-layer graph convolution is performed to learn the representation information of users and items, and the intra-domain and inter-domain high-order representation information of each user and item is obtained. Intra-domain heterogeneous graph convolution: The intra-domain heterogeneous graph only has the nodes of users and items in its own domain. When message aggregation is performed on the nodes, an attention mechanism is used for each node to fuse the representations of its neighbor nodes; Inter-domain heterogeneous graph convolution: The inter-domain heterogeneous graph contains the nodes of users and items in all domains. When message aggregation is performed on the nodes, the nodes in their respective domains among the neighbor nodes are first fused by the intra-domain attention mechanism, and then the representations fused across different domains are fused by the inter-domain attention mechanism, so as to obtain the representation of the nodes. Through the intra-domain and inter-domain heterogeneous graph convolutions, the representations of users and items in the intra-domain and inter-domain are respectively obtained. Through multi-layer convolution, more accurate representation information of user nodes and item nodes is obtained.

[0043] In one embodiment, the method for performing multi-layer graph convolution on the intra-domain heterogeneous graph to obtain the high-order representation information of each user node or item node is as follows: perform three-layer graph convolution on each intra-domain heterogeneous graph, and use a hierarchical attention mechanism for each user node or item node to fuse the representation information of its neighbor nodes for message aggregation, so as to obtain the high-order representation information of all user nodes or item nodes in each intra-domain heterogeneous graph. The method for performing multi-layer graph convolution on the inter-domain heterogeneous graph to obtain the high-order representation information of each user node or item node is as follows: perform three-layer graph convolution on the inter-domain heterogeneous graph, and use a hierarchical attention mechanism for each user node or item node to perform message aggregation, so as to obtain the high-order representation information of all user nodes or item nodes in each inter-domain heterogeneous graph.

[0044] In the embodiment of the present application, graph convolution is respectively performed on the intra-domain and inter-domain heterogeneous graphs to obtain the intra-domain and inter-domain representations of users and items and

[0045] In the intra-domain heterogeneous graph, there is an intra-domain heterogeneous graph for domain a Among them, there is a user representation for the user node There is an item representation for the item node For node i, there are neighbor nodes The graph convolution step is to calculate attention weights, message passing, and message aggregation.

[0046] Calculating attention weights: In the intra-domain heterogeneous graph, for the target node i and its neighbor nodes The importance of its adjacent node j to the target node i, that is, the attention score α, is obtained through a linear transformation ij :[[]]

[0047]

[0048] where a and W are weight parameters, and LeakyReLU is an activation function.

[0049] Message passing: The message passed by the neighbor node j of the target node i:

[0050] Message(j) = Wh j

[0051] Message aggregation: Node i aggregates the representations Message(j) passed by its neighbor nodes.

[0052]

[0053] In this embodiment, through the convolution of three layers of graph neural networks, the final representation of the user or item of the intra-domain node i can be obtained:

[0054]

[0055] In the inter-domain heterogeneous graph, message aggregation is divided into two steps, and two-layer attention mechanisms within the domain and between domains are respectively adopted. For the neighbor nodes of the user node i which include item nodes in three domains a, b, and c and Aggregate the nodes in this domain using the multi-head attention mechanism to obtain and The purpose of the multi-head attention mechanism in this layer is to filter out unimportant node information in each domain. Then calculate the obtained and The computational force weights with node i, and finally fuse to obtain the representation h of node i i . The specific steps of message aggregation in the inter-domain heterogeneous graph are: calculating the attention weights within the domain, message passing, and message aggregation.

[0056] Calculating the attention weights within the domain: In the inter-domain heterogeneous graph, for node i and its neighbor nodes and The attention weights are calculated respectively, and and

[0057]

[0058] where a and W are the weight parameters for each domain, and LeakyReLU is the activation function.

[0059] Message passing: The message passed by the neighbor nodes of the target node i is

[0060]

[0061] Message aggregation: For the target node i, the node aggregation representations of the three domains a, b, and c are calculated respectively and

[0062]

[0063] Therefore, the node representations of domains a, b, and c can be obtained by summing the multi-head attentions within the domain and

[0064]

[0065] Then, the attention weights between domains are calculated to obtain

[0066]

[0067] Finally, the final representations of the inter-domain users and items are obtained

[0068]

[0069] In this embodiment, the intra-domain and inter-domain heterogeneous graphs are convolved through a three-layer graph neural network to obtain the intra-domain user and item representations and where d ∈ {a, b, c}, and the inter-domain user and item representations and

[0070] In one embodiment, the step of fusing the high-order representation information of the user nodes and / or item nodes to obtain the fused representation information of the user nodes and / or item nodes in each domain includes the following steps: Fusing the high-order representation information of all the user nodes in the heterogeneous graph within each domain with the high-order representation information of all the user nodes in the inter-domain heterogeneous graph to obtain the fused representation information of all the user nodes in each domain; Fusing the high-order representation information of all the item nodes in the heterogeneous graph within each domain with the high-order representation information of all the item nodes in the inter-domain heterogeneous graph to obtain the fused representation information of all the item nodes in each domain. The fusion method includes at least one of: concatenation, taking the maximum value, taking the minimum value, taking the average value, and summation.

[0071] As an embodiment of the present application, preferably, the average value method is used for fusion. When the number of domains is m, at this time, the method of using the average value for fusion is:

[0072]

[0073]

[0074] Wherein, is the high-order representation information of the user nodes within the domain, is the high-order representation information of the item nodes within the domain, is the high-order representation information of the inter-domain users, E interi is the high-order representation information of the inter-domain items, n1, n2, …, n m There are a total of m domains n, and are the fused representation information of the user nodes or item nodes respectively. The calculation method of the predicted value of the user's score for the item is:

[0075]

[0076]

[0077]

[0078] …

[0079]

[0080]

[0081] Wherein: n m refers to m domains, d is one of them, is the predicted value of the user's score for the item, and The representation information of the user node and the item node respectively, where is the weight matrix of domain d at the L-th layer, is the bias of domain d at the L-th layer, Relu and Sigmoid are activation functions, and L is the number of fully connected layers, is to the value obtained after splicing and calculating the matrix.

[0082] In the embodiment of the present application, the high-order representation information of the user node or the item node is fused. Taking the number of domains as 3 as an example, for domains a, b, and c, the user and item representations within their respective domains are fused with the user and item representations between domains. Those skilled in the art know that the fusion methods may include splicing, maximum value, minimum value, average value, and summation, etc. After fusion, the user and item representations of domains a, b, and c are obtained respectively. When performing node aggregation in the heterogeneous graph, a hierarchical attention mechanism is adopted to filter important information for each node, thereby alleviating the knowledge negative transfer problem in multi-object cross-domain recommendation and improving the recommendation accuracy.

[0083] In this embodiment, taking the number of domains as three as an example, the user and item representations within the domain obtained in the previous step and where d ∈ {a, b, c} are fused with the user and item representations between domains and The fusion methods include splicing, average value, summation, maximum value, and minimum value, etc. In this embodiment, the average value method is adopted to fuse them to obtain and where d ∈ {a, b, c}.

[0084]

[0085]

[0086] In one embodiment, the fused representation information of all the user nodes or item nodes is respectively input into their respective fully connected neural networks to learn the non-linear relationship between the user and the item, obtain the score prediction value of the user for the item, and perform cross-domain recommendation based on the score prediction value.

[0087] In this embodiment, the non-linear relationship between the user and the item can be obtained in various ways. Preferably, by the method of respectively inputting the fused representation information of all the user nodes or item nodes into their respective fully connected neural networks, the score prediction value of the user for the item is obtained, so that the recommendation is more accurate and the negative transfer of knowledge is significantly alleviated.

[0088] As an embodiment of the present application, taking the number of domains as 3 as an example, the specific method is described. The fully connected layer is used for each of the domains a, b, and c to obtain the final predicted value and

[0089]

[0090]

[0091]

[0092] …

[0093]

[0094]

[0095] where W is the weight matrix, n m refers to m domains, d is one of them, is the predicted value of the user's rating for the item, and are the representation information of the user node and the item node respectively, where is the weight matrix of domain d at the L-th layer, is the bias of domain d at the L-th layer, Relu and Sigmoid are activation functions, and L is the number of layers of the fully connected layer.

[0096] In this embodiment, for each domain, there is an optimization objective:

[0097]

[0098]

[0099] where and are the representations of the user and the item within the domain respectively, and are the representations of the user and the item between domains respectively, Θ is the model parameter, Y + is the positive sample, Y - is the negative sample. Since all the data in the dataset are positive samples, negative sampling is performed on the items that the user has not interacted with and is used to replace Y - . is L2 regularization, where λ is the regularization parameter. The loss function is:

[0100]

[0101] For domains a, b, and c, the total loss function is:

[0102]

[0103] In this embodiment, for each user, their latest interaction record is selected and put into the test set, and the remaining interaction records are put into the training set. Since there are only positive samples in the dataset, in the training set, 4 items are negatively sampled for each user from the items they have not interacted with. In the test set, 99 items are negatively sampled for each user. The representation dimensions of users and items are 32. In the heterogeneous graph convolutional layer, three layers of GCN are used, the multi-head attention mechanism has 3 heads, and the fully connected layer is [128, 64, 32, 16]. Through the above process, the predicted values and Those skilled in the art can select content information that meets the requirements for recommendation based on the predicted values according to actual needs.

[0104] In the embodiments of this application, the high-order relationship between users and items is fully learned through the graph neural network to obtain accurate user and item representations; when performing node aggregation in the heterogeneous graph, a hierarchical attention mechanism is adopted to filter important information for each node, thereby alleviating the problem of knowledge negative transfer in multi-object cross-domain recommendation and improving the recommendation accuracy.

[0105] In one embodiment, as Figure 4 shown, a multi-object cross-domain recommendation device based on a heterogeneous graph is provided. The multi-object cross-domain recommendation device can be integrated into the above computer device 120. The device includes:

[0106] An intra-domain heterogeneous graph and inter-domain heterogeneous graph construction module, used to construct an intra-domain heterogeneous graph and an inter-domain heterogeneous graph based on intra-domain user information and / or item information; a high-order representation information acquisition module, used to perform multi-layer graph convolution on all the intra-domain heterogeneous graphs and inter-domain heterogeneous graphs respectively to obtain high-order representation information of each user node and / or item node; a fused representation information acquisition module, used to fuse the high-order representation information of the user nodes and / or item nodes to obtain fused representation information of user nodes and / or item nodes in each domain; a cross-domain recommendation module, used to respectively input the fused representation information of all the user nodes and item nodes into their respective fully connected neural networks to learn the non-linear relationship between users and items, obtain the score prediction value of users for items, and perform cross-domain recommendation based on the score prediction value.

[0107] It can be understood that the above device includes multiple modules, which can be spliced and combined and can run on various cloud platforms or various devices with computing and processing capabilities. For the description of the operation mode of the above device, reference can be made to the steps of a multi-object cross-domain recommendation method based on a heterogeneous graph, which will not be elaborated here. In the embodiments of the present application, through the construction of intra-domain and inter-domain heterogeneous graphs and graph neural networks, the high-order relationships between users and items are fully learned to obtain accurate user and item representations; when performing node aggregation in the heterogeneous graph, a hierarchical attention mechanism is adopted to filter important information for each node, thereby alleviating the knowledge negative transfer problem in multi-object cross-domain recommendation and improving the recommendation accuracy.

[0108] In one embodiment, a multi-object cross-domain recommendation system based on a heterogeneous graph is provided. When the system runs, it executes the steps of the above-mentioned multi-object cross-domain recommendation method based on a heterogeneous graph.

[0109] In the embodiments of the present application, for the description of the operation mode of the above system, reference can be made to the steps of a multi-object cross-domain recommendation method based on a heterogeneous graph, which will not be elaborated here. In the embodiments of the present application, through the construction of intra-domain and inter-domain heterogeneous graphs and graph neural networks, the high-order relationships between users and items are fully learned to obtain accurate user and item representations; when performing node aggregation in the heterogeneous graph, a hierarchical attention mechanism is adopted to filter important information for each node, thereby alleviating the knowledge negative transfer problem in multi-object cross-domain recommendation and improving the recommendation accuracy.

[0110] It should be understood that although the steps in the flowcharts of the embodiments of the present application are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least a part of other steps or sub-steps or stages of other steps.

[0111] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memories.

[0112] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the various technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0113] The above-described embodiments merely represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A multi-objective cross-domain recommendation method based on heterogeneous graphs, characterized in that The method includes: Constructing an intra-domain heterogeneous graph and an inter-domain heterogeneous graph based on the user information and / or item information within the domain; Performing multi-layer graph convolution on all the intra-domain heterogeneous graphs and inter-domain heterogeneous graphs respectively to obtain the high-order representation information of each user node and / or item node; Fusing the high-order representation information of the user nodes and / or item nodes to obtain the fused representation information of the user nodes and / or the fused representation information of the item nodes in each domain; Inputting the fused representation information of all the user nodes and the fused representation information of the item nodes into their respective fully connected neural networks to learn the non-linear relationship between users and items, obtaining the predicted rating value of the user for the item, and performing cross-domain recommendation based on the predicted rating value; The method for performing multi-layer graph convolution on the intra-domain heterogeneous graph to obtain the high-order representation information of each user node and / or item node is: Performing three-layer graph convolution on each intra-domain heterogeneous graph, and using a hierarchical attention mechanism to fuse the representation information of its neighbor nodes for each user node and / or item node for message aggregation, so as to obtain the high-order representation information of all user nodes and / or item nodes in each intra-domain heterogeneous graph; The method for performing multi-layer graph convolution on the inter-domain heterogeneous graph to obtain the high-order representation information of each user node and / or item node is: Performing three-layer graph convolution on the inter-domain heterogeneous graph, and using a hierarchical attention mechanism for message aggregation for each user node and / or item node, so as to obtain the high-order representation information of all user nodes and / or item nodes in each inter-domain heterogeneous graph; The fusion adopts the average value method, and the calculation method is: ； Among them, is the high-order representation information of the user node within the domain, is the high-order representation information of the item node within the domain, is the high-order representation information of the user across domains, is the high-order representation information of the item across domains, There are a total of m domains n, which are the fusion representation information of the user node or the item node respectively; The calculation method of the predicted rating value of the user for the item is: ； Wherein: represents m domains, , is the predicted value of the user's score for the item, are the representation information of the user node and the item node respectively, where is the weight matrix of domain at the L-th layer, is the bias of domain at the L-th layer, Relu and Sigmoid are activation functions, and L is the number of layers of the fully connected layer.

2. The multi-objective cross-domain recommendation method based on heterogeneous graph according to claim 1, wherein The method for constructing an intra-domain heterogeneous graph and an inter-domain heterogeneous graph based on the user information and / or item information within the domain is: Constructing a corresponding intra-domain heterogeneous graph for each domain, and each intra-domain heterogeneous graph contains at least one user node and / or one item node, the user information corresponds to the user node, and the item information corresponds to the item node; Constructing an inter-domain heterogeneous graph according to the interaction information between the user node and the item node; There is one inter-domain heterogeneous graph.

3. A multi-objective cross-domain recommendation method based on a heterogeneous graph according to claim 1, characterized in that, Fusing the high-order representation information of the user nodes and / or item nodes to obtain the fused representation information of the user nodes and / or the fused representation information of the item nodes in each domain includes the following steps: Fusing the high-order representation information of all the user nodes in each intra-domain heterogeneous graph with the high-order representation information of all the user nodes in the inter-domain heterogeneous graph respectively to obtain the fused representation information of all user nodes in each domain; Fusing the high-order representation information of all the item nodes in each intra-domain heterogeneous graph with the high-order representation information of all the item nodes in the inter-domain heterogeneous graph respectively to obtain the fused representation information of all item nodes in each domain.

4. The multi-objective cross-domain recommendation method based on heterogeneous graph according to claim 3, wherein The fusion methods include at least one of concatenation, taking the maximum value, taking the minimum value, average value, and summation.

5. A multi-objective cross-domain recommendation device based on a heterogeneous graph, characterized in that, The device includes: An intra-domain heterogeneous graph and inter-domain heterogeneous graph construction module, configured to construct an intra-domain heterogeneous graph and an inter-domain heterogeneous graph based on the user information and / or item information within the domain; The high-order representation information acquisition module is used to perform multi-layer graph convolution on all the intra-domain heterogeneous graphs and inter-domain heterogeneous graphs respectively to obtain the high-order representation information of each user node and / or item node; The fused representation information acquisition module is used to fuse the high-order representation information of the user node and / or item node to obtain the fused representation information of the user node and / or the fused representation information of the item node in each domain; The cross-domain recommendation module is used to separately input the fused representation information of all the user nodes and the fused representation information of the item nodes into their respective fully connected neural networks to learn the non-linear relationship between the user and the item, obtain the predicted rating value of the user for the item, and perform cross-domain recommendation based on the predicted rating value; The method for obtaining the high-order representation information of each user node and / or item node by performing multi-layer graph convolution on the intra-domain heterogeneous graph is as follows: Perform three-layer graph convolution on each intra-domain heterogeneous graph, and use a hierarchical attention mechanism for each user node and / or item node to fuse the representation information of its neighbor nodes for message aggregation, so as to obtain the high-order representation information of all user nodes and / or item nodes in each intra-domain heterogeneous graph; The method for obtaining the high-order representation information of each user node and / or item node by performing multi-layer graph convolution on the inter-domain heterogeneous graph is as follows: Perform three-layer graph convolution on the inter-domain heterogeneous graph, and use a hierarchical attention mechanism for each user node and / or item node for message aggregation, so as to obtain the high-order representation information of all user nodes and / or item nodes in each inter-domain heterogeneous graph; The fusion adopts the average value method, and the calculation method is as follows: ； Among them, is the high-order representation information of the in-domain user node, is the high-order representation information of the in-domain item node, is the high-order representation information of the inter-domain user, is the high-order representation information of the inter-domain item, There are a total of m domains n, which are respectively the fusion representation information of the user node or the item node; The calculation method of the predicted rating value of the user for the item is as follows: ； Wherein: represents m domains, , is the predicted value of the user's rating of the item, are the representation information of the user node and the item node respectively, where is the weight matrix of domain at the L-th layer, is the bias of domain at the L-th layer, Relu and Sigmoid are activation functions, and L is the number of layers of the fully connected layer.

6. A multi-objective cross-domain recommendation system based on a heterogeneous graph, characterized in that When the system is running, it executes the steps of a multi-objective cross-domain recommendation method based on a heterogeneous graph as described in any one of claims 1 to 4.

Images