A cross-domain product recommendation method based on dual-view fusion heterogeneous graph neural network

Abstract

The present invention discloses a cross-domain commodity recommendation method based on a dual-perspective fusion heterogeneous graph neural network, including: reading a user and commodity shopping database, constructing a node graph of a heterogeneous shopping network, and converting the node graph into a line graph according to the edge connection in the node graph; obtaining a semantic embedding vector and a structural embedding vector from two perspectives, namely, a meta-path and a biased bidirectional random walk, in the node graph, and then obtaining a node pair embedding vector of a user-commodity; obtaining an edge embedding vector of a user-commodity according to the line graph using an edge embedding algorithm; obtaining a mapping function from a node domain to an edge domain using a meta-network according to the node pair embedding vector and the edge embedding vector; using a mapping function, mapping the node pair embedding vector of a commodity and a target user to an edge domain, and recommending commodities to the target user according to the mapped edge embedding vector. The present invention can effectively utilize node and interaction information and improve the accuracy of commodity recommendation.

Description

Technical Field

[0001] The present invention relates to the field of data analysis technology, and in particular to a cross-domain product recommendation method based on dual-view fusion heterogeneous graph neural network. Background Art

[0002] Various shopping scenarios in the real world can be accurately modeled into heterogeneous shopping networks composed of multiple types of nodes and edges. In this heterogeneous shopping network, users and products are nodes, and interactions are edges. Heterogeneous graph neural networks can effectively retain user preferences and product features using this heterogeneous structure, provide more accurate product recommendations, thereby enhancing user shopping experience and improving user stickiness, creating economic benefits for enterprises.

[0003] The product recommendation accuracy of the existing meta-path-based heterogeneous graph neural network is not high, and there are two problems in using the node similarity between users and products for recommendation: on the one hand, aggregating the node information of users and products based solely on meta-paths may lead to information loss or redundancy. For example, if "user-product-user" and "product-user-product" are selected as meta-paths, users can only aggregate node information on low-order neighbors, and may lose user preference information at a longer distance. If longer meta-paths are added, a large amount of node information will be repeated, resulting in redundancy. On the other hand, directly using the node similarity between users and products for recommendation is not very interpretable and will lose information about the interactive relationship between users and products.

[0004] Therefore, how to effectively utilize nodes and interaction information to improve the accuracy of product recommendations is still a problem that needs to be solved urgently by those skilled in the art. Summary of the invention

[0005] The purpose of the present invention is to provide a cross-domain product recommendation method based on dual-view fusion heterogeneous graph neural network to effectively utilize node and interaction information and improve the accuracy of product recommendation.

[0006] A cross-domain product recommendation method based on dual-view fusion heterogeneous graph neural network, including steps S1 to S5:

[0007] Step S1, reading the user and product shopping database, constructing a node graph of a heterogeneous shopping network, and converting the node graph into a line graph according to the edge conditions in the node graph;

[0008] Step S2, obtaining a semantic embedding vector and a structural embedding vector from the two perspectives of meta-path and biased bidirectional random walk in the node graph, and obtaining a user-item node pair embedding vector based on the semantic embedding vector and the structural embedding vector;

[0009] Step S3, according to the line graph, using the edge embedding algorithm to obtain the user-item edge embedding vector;

[0010] Step S4, obtaining a mapping function from a node domain to an edge domain using a meta-network according to a user-item node pair embedding vector and a user-item edge embedding vector;

[0011] Step S5: Map the node pair embedding vectors of the product and the target user to the edge domain using a mapping function, and recommend products to the target user based on the mapped edge embedding vectors.

[0012] The cross-domain product recommendation method based on dual-view fusion heterogeneous graph neural network provided by the present invention has the following beneficial effects:

[0013] (1) The modeling information loss of the present invention is small

[0014] Traditional modeling methods model the commodity shopping network as a homogeneous network, ignoring the heterogeneity of node types and relationships between nodes, resulting in large modeling information loss. However, this method reads the interaction dataset of users and commodities from the shopping database of users and commodities, constructs the node graph of the heterogeneous shopping network, and retains rich semantic information;

[0015] (2) The present invention has rich node representations of users and products

[0016] Traditional recommendation systems mostly rely on meta-paths for information extraction, lack information supplementation outside the meta-path, and rarely learn complex structural information in heterogeneous commodity shopping networks, resulting in insufficient node representation. The present invention integrates the semantic embedding vector based on the meta-path perspective and the structural embedding vector based on the biased bidirectional random walk perspective to obtain the user-commodity node pair embedding vector. Through the dual perspective, more node information can be captured to enhance the representation ability of users and commodities, and obtain richer node representations of users and commodities.

[0017] (3) The product recommendation of the present invention is highly interpretable and has high product recommendation accuracy

[0018] Existing recommendation systems usually make recommendations based on the similarity of user-item node pairs, ignoring the side information between users and items. This method maps the user-item node pair embedding vector to the edge domain through a meta-network, and uses the mapped edge embedding vector to recommend items. This not only captures the user-item side information, improves the accuracy of item recommendations, but also makes them more interpretable. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the cross-domain product recommendation method based on dual-perspective fusion heterogeneous graph neural network of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] See also Figure 1 The cross-domain product recommendation method based on dual-view fusion heterogeneous graph neural network provided by the embodiment of the present invention includes steps S1 to S5.

[0022] Step S1, read the user and product shopping database, construct a node graph of the heterogeneous shopping network, and convert the node graph into a line graph according to the edge conditions in the node graph.

[0023] Wherein, step S1 specifically includes:

[0024] Step S11, read the user and product interaction data set from the user and product shopping database, and construct a node graph of the heterogeneous shopping network G , G =( V , E , T V , T E ),in, V is a collection of nodes, V Nodes in v Indicates a user or a product. E is the set of edges, E The edge e Represents the relationship between nodes. T V is a collection of node types, T V ={ t 1, t 2}, t 1 indicates the user node type, t 2 indicates the commodity node type, T E It is a collection of edge types, representing different types of relationships between nodes, such as the purchase relationship between users and products, the attention relationship between users, and the brand relationship between products. G After that, set the meta path collection ;

[0025] Step S12: GConvert to line graph , Each node in the graph corresponds to G One edge in The expression is:

[0026] ;

[0027] in, V L is the node set of the line graph, , Nodes in yes G One edge in v i , v j It is a node graph G Nodes in i , j , E L is the edge set of the line graph, , v s It is a node graph G Nodes in s ,like G The edge and All v j As a vertex, and exist Connected in.

[0028] Step S2: Obtain semantic embedding vectors and structural embedding vectors from the perspectives of meta-path and biased bidirectional random walk in the node graph, and obtain user-item node pair embedding vectors based on the semantic embedding vectors and structural embedding vectors.

[0029] Wherein, step S2 specifically includes:

[0030] Step S21, in the node graph, by SVD feature Method selection The target meta-path instance in ;

[0031] Step S22: Aggregate the target meta-path instances through the attention mechanism to obtain the semantic embedding vector of the node ;

[0032] Step S23, aggregate the biased bidirectional random walk sequence through the attention mechanism to obtain the structural embedding vector of the node ;

[0033] Step S24, embedding the semantics into vector and structure embedding vector Concatenate to get the node embedding vector , and then get the user's node embedding vector Node embedding vector of the product ;

[0034] Step S25: embedding the user's node vector Node embedding vector of the product Concatenate to obtain the user-item node pair embedding vector .

[0035] Wherein, step S21 specifically includes:

[0036] The meta-path instances are measured by calculating the similarity of the nodes between the meta-path instances. The similarity calculation formula is:

[0037] ;

[0038] in, Represents the similarity between nodes. SVD feature It represents a method for obtaining node similarity through singular value decomposition, which can extract and utilize potential features in data, thereby improving the performance of capturing similarity between nodes;

[0039] For a meta path instance ,in, v 1. v 2. v m It is a node graph G The first, second, and m nodes, m Represents a meta path instance P The total number of nodes in the path, calculate the similarity between adjacent nodes on the path, and take the average to get the average similarity :

[0040] ;

[0041] in, v i+1 Represents a meta path instance Path middle v i The next node of

[0042] For the meta path collection , sort the meta-path instances under each meta-path from large to small according to the average similarity, and select the first K meta-path instance as the target meta-path instance;

[0043] Step S22 specifically includes:

[0044] The sequence of target meta-path instances is embedded into a low-dimensional space through a convolutional neural network to obtain the semantic embedding vector of the meta-path instance. , the expression is:

[0045] ;

[0046] in, CNN represents a convolutional neural network, represents the encoding matrix of the target meta-path instance, represents all relevant parameters in the convolutional neural network that processes the sequence of meta-paths;

[0047] The attention mechanism is used to aggregate instances under the same target meta-path to obtain the embedding vectors of users and products under the target meta-path. The expression is as follows:

[0048] ;

[0049] ;

[0050] ;

[0051] in, Represents the weighted summation of r The embedding vectors of users and products under the meta-path of the type, avg represents the average operation, Indicates that The attention mechanism of each head is concatenated to calculate the The weight of each meta-path instance, Indicates The embedding vector of each meta-path instance, Indicates Metapath instance pair node v The importance score of LeakyReLU represents the activation function used to increase the nonlinear characteristics of the neural network. represents the parameter matrix of the feedforward neural network that processes the sequence of meta-paths;

[0052] The attention mechanism is used to aggregate the embedding vectors of users and products under different types of meta-paths to obtain the semantic embedding vector of the node. , the expression is:

[0053] ;

[0054] ;

[0055] ;

[0056] in, Expressed as r The weight of each type of meta-path, Indicates r Type of meta-path pair node v The importance score of A feature vector representing the attention parameters;

[0057] Step S23 specifically includes:

[0058] Perform biased bidirectional random walks and record the node sequences of users and products that the forward and reverse walks pass through. The walk transition probability matrix is ​​defined as:

[0059] ;

[0060] in, represents the transition probability matrix of the node, , , They are respectively assigned to the first hop, the second hop, and the k The weight of the neighbor to jump to, , , It is the node step 1, step 2, k The probability transfer matrix of the step.

[0061] Elements in It is a slave node i To Node j of k Step transition probability, where the transition probability for:

[0062] ;

[0063] in, Indicates included in v nodes in all meta-path instances of Indicates absence v Node in the meta path instance of . Ability to describe nodes and nodes The similarity between . and They represent the process of jumping from the current node to the node in the random walk process. and nodes The probability of , to prioritize sampling those nodes that are not on the target node v A node in the meta-path instance of .

[0064] The biased bidirectional random walk sequence is embedded into a low-dimensional space using a convolutional neural network to obtain the structural embedding vector of the meta-path instance. , the expression is:

[0065] ;

[0066] in, Indicates The encoding matrix of the subsampled node sequence, represents all relevant parameters in the convolutional neural network that processes the random walk sequence;

[0067] Use the multi-head attention mechanism to assign weights to different walk sequences, analyze the contribution of each sequence to users and products, and obtain the structural embedding vector of the node , the expression is:

[0068] ;

[0069] ;

[0070] in, Indicates The attention mechanism of each head calculates the The weight of a walk sequence, represents the parameter matrix of the feedforward neural network processing the random walk sequence;

[0071] Step S24 satisfies the following conditional formula:

[0072] ;

[0073] in, Represents a splicing operation;

[0074] Step S25 satisfies the following conditional formula:

[0075] .

[0076] Step S3: According to the line graph, an edge embedding algorithm is used to obtain the user-item edge embedding vector.

[0077] Wherein, step S3 satisfies the following conditional formula:

[0078] ;

[0079] in, represents the edge embedding vector of user-item, line 2 vec It is an edge embedding algorithm for line graphs.

[0080] Step S4: According to the user-item node pair embedding vector and the user-item edge embedding vector, a mapping function from the node domain to the edge domain is obtained using the meta-network.

[0081] Wherein, step S4 specifically includes:

[0082] Embedding vectors based on user-item edges through meta-network Train to obtain the mapping function from node domain to edge domain , mapping function Parameters Satisfy the following formula:

[0083] ;

[0084] in, represents the meta-network, are the parameters in the meta-network.

[0085] Step S5: Map the node pair embedding vectors of the product and the target user to the edge domain using a mapping function, and recommend products to the target user based on the mapped edge embedding vectors.

[0086] Wherein, step S5 specifically includes:

[0087] Step S51: embed the node pair of the product and the target user into a vector Input to the mapping function , get the edge embedding vector after mapping , the expression is:

[0088] ;

[0089] Step S52: embed the mapped edge into vector Input to the multi-layer perceptron to obtain the recommendation score ,according to Perform product recommendation tasks for target users, such as selecting several products with recommendation scores greater than a threshold and recommending them to target users. The expression is:

[0090] ;

[0091] in, MLP represents a multi-layer perceptron;

[0092] The loss function L of the product recommendation task is:

[0093] ;

[0094] in, represents the minimum value, R represents the edge between the target user and the product in the node domain, Is a user u For goods The real rating.

[0095] In summary, the cross-domain product recommendation method based on dual-view fusion heterogeneous graph neural network provided by the present invention has the following beneficial effects:

[0096] (1) The modeling information loss of the present invention is small

[0097] Traditional modeling methods model the commodity shopping network as a homogeneous network, ignoring the heterogeneity of node types and relationships between nodes, resulting in large modeling information loss. However, this method reads the interaction dataset of users and commodities from the shopping database of users and commodities, constructs the node graph of the heterogeneous shopping network, and retains rich semantic information;

[0098] (2) The present invention has rich node representations of users and products

[0099] Traditional recommendation systems mostly rely on meta-paths for information extraction, lack information supplementation outside the meta-path, and rarely learn complex structural information in heterogeneous commodity shopping networks, resulting in insufficient node representation. The present invention integrates the semantic embedding vector based on the meta-path perspective and the structural embedding vector based on the biased bidirectional random walk perspective to obtain the user-commodity node pair embedding vector. Through the dual perspective, more node information can be captured to enhance the representation ability of users and commodities, and obtain richer node representations of users and commodities.

[0100] (3) The product recommendation of the present invention is highly interpretable and has high product recommendation accuracy

[0101] Existing recommendation systems usually make recommendations based on the similarity of user-item node pairs, ignoring the side information between users and items. This method maps the user-item node pair embedding vector to the edge domain through a meta-network, and uses the mapped edge embedding vector to recommend items. This not only captures the user-item side information, improves the accuracy of item recommendations, but also makes them more interpretable.

[0102] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0103] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A cross-domain product recommendation method based on dual-view fusion heterogeneous graph neural network, characterized in that: It includes steps S1 to S5: Step S1, reading the user and product shopping database, constructing a node graph of a heterogeneous shopping network, and converting the node graph into a line graph according to the edge conditions in the node graph; Step S2, obtaining a semantic embedding vector and a structural embedding vector from the two perspectives of meta-path and biased bidirectional random walk in the node graph, and obtaining a user-item node pair embedding vector based on the semantic embedding vector and the structural embedding vector; Step S3, according to the line graph, using the edge embedding algorithm to obtain the user-item edge embedding vector; Step S4, obtaining a mapping function from a node domain to an edge domain using a meta-network according to a user-item node pair embedding vector and a user-item edge embedding vector; Step S5, using a mapping function, mapping the node pair embedding vectors of the product and the target user to the edge domain, and recommending products to the target user based on the mapped edge embedding vectors; Step S2 specifically includes: Step S21, in the node graph, by SVD feature Method selection The target meta-path instance in SVD feature It represents the method of obtaining node similarity through singular value decomposition; Step S22: Aggregate the target meta-path instances through the attention mechanism to obtain the semantic embedding vector of the node ; Step S23, aggregate the biased bidirectional random walk sequence through the attention mechanism to obtain the structural embedding vector of the node ; Step S24, embedding the semantics into vector and structure embedding vector Concatenate to get the node embedding vector , and then get the user's node embedding vector Node embedding vector of the product ; Step S25: embedding the user's node vector Node embedding vector of the product Concatenate to obtain the user-item node pair embedding vector .

2. The cross-domain product recommendation method based on dual-view fusion heterogeneous graph neural network according to claim 1 is characterized in that: Step S1 specifically includes: Step S11, read the user and product interaction data set from the user and product shopping database, and construct a node graph of the heterogeneous shopping network G , G =( V , E , T V , T E ),in, V is a collection of nodes, V Nodes in v Indicates a user or a product. E is the set of edges, E The edge e Represents the relationship between nodes. T V is a collection of node types, T V ={ t 1, t 2}, t 1 indicates the user node type, t 2 indicates the commodity node type, T E is a set of edge types, and we get the node graph G After that, set the meta path collection ; Step S12: G Convert to line graph , Each node in the graph corresponds to G One edge in The expression is: ； in, V L is the node set of the line graph, , Nodes in yes G One edge in v i , v j It is a node graph G The i Nodes i , No. j nodes, E L is the edge set of the line graph, , v s It is a node graph G The s nodes, if G The edge and All v j As a vertex, and exist Connected in.

3. The cross-domain product recommendation method based on dual-view fusion heterogeneous graph neural network according to claim 1 is characterized by: Step S21 specifically includes: The meta-path instances are measured by calculating the similarity of the nodes between the meta-path instances. The similarity calculation formula is: ； in, Indicates the similarity between nodes; For a meta path instance ,in, v 1. v 2. v m It is a node graph G The first, second, and m nodes, m Represents a meta path instance P The total number of nodes in the path is calculated, and the similarity between adjacent nodes on the path is calculated and the average is taken to obtain the average similarity. : ； in, v i+1 Represents a meta path instance Path middle v i The next node of For the meta path collection , sort the meta-path instances under each meta-path from large to small according to the average similarity, and select the first K meta-path instance as the target meta-path instance; Step S22 specifically includes: The sequence of target meta-path instances is embedded into a low-dimensional space through a convolutional neural network to obtain the semantic embedding vector of the meta-path instance. , the expression is: ； in, CNN represents a convolutional neural network, represents the encoding matrix of the target meta-path instance, represents all relevant parameters in the convolutional neural network that processes the sequence of meta-paths; The attention mechanism is used to aggregate instances under the same target meta-path to obtain the embedding vectors of users and products under the target meta-path. The expression is as follows: ； ； ； in, Represents the weighted sum of r The embedding vectors of users and products under the meta-path of the type, avg represents the average operation, Indicates that The attention mechanism of each head is concatenated to calculate the The weight of each meta-path instance, Indicates The embedding vector of each meta-path instance, Indicates Metapath instance pair node v The importance score of LeakyReLU represents the activation function used to increase the nonlinear characteristics of the neural network. represents the parameter matrix of the feedforward neural network that processes the sequence of meta-paths; The attention mechanism is used to aggregate the embedding vectors of users and products under different types of meta-paths to obtain the semantic embedding vector of the node. , the expression is: ； ； ； in, Expressed as r The weight of each type of meta-path, Indicates r Type of meta-path pair node v The importance score of A feature vector representing the attention parameters; Step S23 specifically includes: Perform biased bidirectional random walks and record the node sequences of users and products that the forward and reverse walks pass through. The walk transition probability matrix is ​​defined as: ； in, represents the transition probability matrix of the node, , , They are respectively assigned to the first hop, the second hop, and the k The weight of the neighbor to jump to, , , It is the node step 1, step 2, k The probability transfer matrix of the step; The biased bidirectional random walk sequence is embedded into a low-dimensional space using a convolutional neural network to obtain the structural embedding vector of the meta-path instance. , the expression is: ； in, Indicates The encoding matrix of the subsampled node sequence, represents all relevant parameters in the convolutional neural network that processes the random walk sequence; Use the multi-head attention mechanism to assign weights to different walk sequences, analyze the contribution of each sequence to users and products, and obtain the structural embedding vector of the node , the expression is: ； ； in, Indicates The attention mechanism of each head calculates the The weight of a walk sequence, represents the parameter matrix of the feedforward neural network processing the random walk sequence; Step S24 satisfies the following conditional formula: ； in, Represents a splicing operation; Step S25 satisfies the following conditional formula: 。 4. The cross-domain product recommendation method based on dual-view fusion heterogeneous graph neural network according to claim 3 is characterized in that: Step S3 satisfies the following conditional formula: ； in, represents the edge embedding vector of user-item, line 2 vec It is an edge embedding algorithm for line graphs.

5. The cross-domain product recommendation method based on dual-view fusion heterogeneous graph neural network according to claim 4 is characterized in that: Step S4 specifically includes: Embedding vectors based on user-item edges through meta-network Train to obtain the mapping function from node domain to edge domain , mapping function Parameters Satisfy the following formula: ； in, represents the meta-network, are the parameters in the meta-network.

6. The cross-domain product recommendation method based on dual-view fusion heterogeneous graph neural network according to claim 5 is characterized in that: Step S5 specifically includes: Step S51: embed the node pair of the product and the target user into a vector Input to the mapping function , get the edge embedding vector after mapping , the expression is: ； Step S52: embed the mapped edge into vector Input to the multi-layer perceptron to obtain the recommendation score ,according to Perform product recommendation tasks for target users. The expression is: ； in, MLP represents a multi-layer perceptron; The loss function L of the product recommendation task is: ； in, represents the minimum value, R represents the edge between the target user and the product in the node domain, represents the number of edges between target users and products in the node domain, Is a user u Give goods The real rating.

Images