Product recommendation method based on heterogeneous graph neural network with parallel feature extraction

Abstract

The present invention discloses a product recommendation method based on a heterogeneous graph neural network with parallel feature extraction. Products are used as nodes, user behaviors are used as edges, and different behaviors are set with different edge weight values to construct a product relationship graph, highlighting user dimension features. The heterogeneous graph neural network with parallel feature extraction includes an internal feature extraction module, an external feature extraction module and a meta-path module. The internal feature extraction module learns the features of nodes of the same type through an attention mechanism and samples strongly correlated nodes. The external feature extraction module captures the dependencies between long-distance nodes in the network topology structure through feature space and topological space, combined with the attention mechanism, and enhances the ability to extract multi-dimensional information propagation features between different types of nodes. The meta-path module extracts deep semantic information. A heterogeneous graph neural network with parallel feature extraction is used to optimize the learning performance of feature embedding of the heterogeneous graph neural network, and solve problems such as insufficient feature extraction capabilities of existing models and inability to capture long-distance dependencies of nodes.

Description

Technical Field

[0001] The present invention belongs to the field of heterogeneous graph node classification recommendation, and specifically relates to a product recommendation method based on a heterogeneous graph neural network with parallel feature extraction. Background Art

[0002] In recent years, graph neural networks (GNNs) have gained popularity in various graph-structured tasks, such as node classification and link prediction, due to their ability to mine and analyze node features and semantic information in graph-structured data (non-Euclidean space) and their superior performance and wide applicability compared to recurrent neural networks (RNNs) and convolutional neural networks (CNNs). However, in the era of big data, the information that needs to be processed in the real world is increasingly interactive, often involving a large number of intertwined objects of different types. For example, in social networks and e-commerce shopping, heterogeneous graph structures contain a variety of objects and relationships. While graph neural network models based on homogeneous networks are currently quite mature, they cannot fully extract multi-type node features and inter-node relationships, hindering their ability to mine the rich semantic information contained in heterogeneous graphs. Consequently, graph neural networks based on heterogeneous graphs have been proposed, which far outperform homogeneous graph networks in information fusion and extracting semantic information from multi-type and multi-relational structures.

[0003] When it comes to product recommendations, existing recommendation models typically focus on product attributes and the product category they belong to. However, in real-world e-commerce shopping, user behavior should be the driving force. Categorizing and recommending products based on clicks, add-to-carts, and order placements can significantly improve user experience and product purchase rates. Therefore, to address the current challenges in product recommendation, it is necessary to develop a heterogeneous graph neural network product recommendation model that deeply mines user behavior and extracts multi-dimensional product node information. Summary of the Invention

[0004] In response to the shortcomings of existing methods, the present invention proposes a heterogeneous graph neural network product recommendation method with parallel feature extraction.

[0005] In a first aspect, the present invention provides a product recommendation method based on a heterogeneous graph neural network with parallel feature extraction, the method comprising the following steps:

[0006] Step (1): Obtain e-commerce product data, including product attributes and user behavior;

[0007] Step (2): Preprocess the e-commerce product data; specifically:

[0008] A product relationship graph G is constructed based on e-commerce product data, wherein products are nodes, product attributes are node attributes, and edges are user behaviors. The product relationship graph G is represented by a feature matrix X. An element value of 0 in the feature matrix X indicates that there is no direct connection between two product nodes, an element value greater than 0 indicates a direct connection between the two product nodes, and an element value of 1 corresponds to a user behavior that the user clicked on the two products within a period of time, an element value of 2 corresponds to a user behavior that the user added the two products to the shopping cart at the same time, and an element value of 3 corresponds to a user behavior that the user ordered the two products at the same time.

[0009] Based on user behavior, extract meta-paths for each node in the product relationship graph G;

[0010] Step (3): Perform initial type classification on the feature matrix X:

[0011] According to the original commodity classification in e-commerce, the same type of nodes and the original edge relationships of these nodes in the feature matrix X are extracted from the feature matrix X to form the same type feature matrix X1;

[0012] Step (4): Use the internal feature module to extract the feature X between nodes of the same type N ;

[0013] Step (5): Using the external feature module, the random attention mechanism is combined with the network topology space and the node feature space to learn the long-range dependency relationship and extract the features E between different types of nodes;

[0014] Step (6): Use the meta-path module to mine the high-level semantic information in the meta-path, further update the network topology, and extract the relationship matrix X M ;

[0015] Step (7): Use the graph convolution module to calculate the feature X between nodes of the same type. N , Features E between different types of nodes, relationship matrix X M , output classification results;

[0016] Step (8): Recommend products based on the product node classification results.

[0017] In a second aspect, the present invention provides a product recommendation system, characterized by comprising:

[0018] Data acquisition module, used to obtain e-commerce product data, including product attributes and user behavior;

[0019] The data preprocessing module is used to preprocess the e-commerce commodity data to obtain the commodity relationship graph G, feature matrix X, and meta-path;

[0020] The same type data extraction module is used to extract the same type of nodes and the original edge relationships of these nodes in the feature matrix X from the feature matrix X to form the same type feature matrix X1;

[0021] Internal feature module, extracting features X between nodes of the same type N ;

[0022] The external feature module uses a random attention mechanism to combine the network topology space and the node feature space to learn long-range dependencies and extract features E between different types of nodes;

[0023] Meta-path module, mining high-level semantic information in meta-path, further updating network topology, extracting relationship matrix X M ;

[0024] The graph convolution module, based on the features X between nodes of the same type N , Features E between different types of nodes, relationship matrix X M , output the classification results.

[0025] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method described above.

[0026] In a fourth aspect, the present invention provides a computing device comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method described is implemented.

[0027] Beneficial effects of the present invention:

[0028] The present invention proposes a heterogeneous graph neural network product recommendation method with parallel feature extraction, using products as nodes and user behaviors as edges, and setting different edge weights for different behaviors to construct a product relationship graph, highlighting user dimension features. The classification of products is carried out in parallel from three parts: feature extraction of nodes of the same type, capturing the dependency between distant nodes in the original structure from the node feature space and the network topology space, and enhancing the model's ability to extract multi-dimensional information propagation features between different types of nodes in the heterogeneous graph. The meta-path is combined with the network topology structure, and an attention mechanism is adopted. A gate mechanism is used in message propagation to eliminate the influence of noise in the original graph on the topology structure. Multiple meta-path information is aggregated to obtain a more accurate meta-path-based network topology structure that can reveal rich semantic information in the network. Finally, the outputs of the three modules are subjected to graph convolution classification to obtain the product classification results. According to the classification results, the edge values of the same category are accumulated to judge the user's interest in the product, improve the accuracy of product recommendations, and enhance user experience and product purchase rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of the process of the present invention.

[0030] Figure 2 Schematic diagram of the heterogeneous graph neural network model structure for parallel feature extraction in the present invention.

[0031] Figure 3 This is a structural diagram of the internal feature extraction module of the present invention.

[0032] Figure 4 This is a structural diagram of the external feature extraction module of the present invention.

[0033] Figure 5 The network topology space in the external feature extraction module of the present invention is combined with the node feature space to generate a node embedded structure diagram.

[0034] Figure 6 It is a structural diagram of the meta-path module of the present invention. DETAILED DESCRIPTION

[0035] The present invention is further described below with reference to the accompanying drawings and specific implementation steps:

[0036] The technical solution adopted by the present invention is that the parallel feature extraction heterogeneous graph neural network (PHGC) model consists of three parts: an internal feature extraction module, an external feature extraction module, and a meta-path module. The internal feature extraction module uses an attention mechanism to learn the features of nodes of the same type, filtering out a large number of weakly correlated nodes and sampling strongly correlated nodes. The external feature extraction module uses two perspectives: feature space and topological space, and combines the attention mechanism to capture the dependencies between distant nodes in the network topology, enhancing the model's ability to extract multidimensional information propagation features between different types of nodes in the heterogeneous graph. Finally, the output is a node embedding that better reflects node features and network topology information. The meta-path module further extracts deep semantic information from the entire heterogeneous graph neural network and, combined with the original network topology, obtains a network topology that better reveals node domains and semantic information. Finally, the outputs of the three parts of the model are input into the graph convolution module for node classification. Based on the product classification results, the edges connecting nodes of the same category are accumulated. The accumulated value is used to determine user interest in the product and make product recommendations.

[0037] A product recommendation method based on heterogeneous graph neural network with parallel feature extraction, such as Figure 1 The following steps are involved:

[0038] Step (1): Obtain e-commerce product data, including product attributes and user behavior;

[0039] Step (2): Preprocess the e-commerce product data; specifically:

[0040] A product relationship graph G is constructed based on e-commerce product data, where products are nodes, product attributes are node attributes, and edges are user behaviors. The product relationship graph G is represented by a feature matrix X. An element value of 0 in the feature matrix X indicates that there is no direct connection between two product nodes, an element value greater than 0 indicates a direct connection between the two product nodes, and an element value of 1 corresponds to a user behavior of the user clicking on the two products within a period of time, an element value of 2 corresponds to a user behavior of the user adding the two products to the shopping cart at the same time, and an element value of 3 corresponds to a user behavior of the user ordering the two products at the same time.

[0041] Based on user behavior, we extract meta-paths from each node in the product relationship graph G. Based on the product relationship graph, we select different meta-paths based on domain experience as higher-level data structures to provide richer input information for feature extraction. For example, if a user purchases product 1, which is produced by a certain manufacturer (a product attribute), then the potential relationship is that the user may also be interested in another product 2 from the same manufacturer (although it may not be of the same type). This meta-path is abstracted into: product 1 - manufacturer - product 2.

[0042] like Figure 2 As shown in Figure 1, the heterogeneous graph neural network framework model for parallel feature extraction consists of three parts: an internal feature extraction module, an external feature extraction module, and a meta-path module. The preprocessed data is input into the three modules for parallel execution.

[0043] Step (3): Perform initial type classification on the feature matrix X:

[0044] According to the original commodity classification in e-commerce, the same type of nodes and the original edge relationships of these nodes in the feature matrix X are extracted from the feature matrix X to form the same type feature matrix X1;

[0045] Step (4): Use the internal feature module to extract features between nodes of the same type;

[0046] like Figure 3 As shown in the figure, the internal feature module first calculates the contribution value of the same type of nodes to the node, that is, its importance, and then selects a threshold and compares the contribution value of the node with the threshold to filter out weakly correlated nodes with low contribution values and collect strongly correlated nodes with high contribution values.

[0047] 4-1 The internal feature module learns the features of nodes of the same type through the attention mechanism and updates the node features:

[0048] Given a pair of nodes v1 and v2 of the same type, the attention mechanism learns the correlation between nodes v1 and v2. That is, the contribution of node v1 to node v2; The calculation formula is as follows:

[0049]

[0050] in are the embedding vectors of nodes v1 and v2 respectively, and att() represents the attention mechanism;

[0051] 4-2 pairs Normalize the nodes using the SoftMax function to get the weight of the nodes. The formula is as follows:

[0052]

[0053] Where i represents the number of nodes; exp() represents the normalization function;

[0054] 4-3 weight values of nodes Compare with the threshold, that is, judge the correlation between two nodes of the same type. If If the correlation is greater than the threshold, the two nodes of the same type are considered to be strongly correlated and no processing is performed. Otherwise, the two nodes of the same type are considered to be weakly correlated and the corresponding elements in the feature matrix X1 of the same type are updated to 0.

[0055] 4-4 Output the updated feature matrix X of the same type through a fully connected layer N ;

[0056] Step (5): Use the external feature module to extract features between different types of nodes;

[0057] like Figure 4 As shown in the figure, the external feature module combines two perspectives: node feature space and network topology space, to solve the long-distance spatial dependence of nodes, explore the connections between different types of nodes in heterogeneous networks, deeply explore the relationship between strongly correlated nodes, and reduce noise interference.

[0058] like Figure 5 As shown in Figure 3, when extracting features from different types of nodes, the external feature module uses a random attention mechanism to combine the network topology space and the node feature space to learn long-range dependencies.

[0059] The external feature module adopts random attention mechanism to combine network topology space and node feature space to learn medium and long distance dependencies;

[0060] 5-1 uses the random attention mechanism combined with the node feature space to learn long-distance dependencies. The specific process is as follows:

[0061] 5-1-1 In the node feature space, after node feature mapping, according to the node feature matrix X and node similarity cossim , take the k nodes with the highest similarity to build the graph GF, whose adjacency matrix is A F ;

[0062] The node similarity cos sim The cosine similarity is calculated as follows:

[0063]

[0064] where X i and X j Represents the adjacency matrix A respectively F The feature vectors of nodes i and j;

[0065] 5-1-2 Feature matrix X based on adjacency matrix A F Continuously update to get node embedding information E F Specifically:

[0066] 5-1-2-1 Initialization, iteration number m = 0, node embedding information

[0067] 5-1-2-2 Use activation function to obtain node embedding information E F , the formula is as follows:

[0068]

[0069] Where Relu represents the activation function, represents the weight matrix, Perform symmetric normalization processing, Indicates the update of the feature matrix X after the mth iteration;

[0070] 5-1-2-3 Judgment and Are they the same? If so, output the final node embedding information. If not, update m=m+1 and return to step 5-1-2-2;

[0071] 5-2 Using random attention mechanism combined with network topology space to learn long-distance dependencies The specific process is as follows:

[0072] 5-2-1 In the network topology space, establish the network topology structure G based on the characteristic matrix X T , whose adjacency matrix is A T ;

[0073] 5-2-2 The feature matrix X is based on the adjacency matrix A T Continuously update to get node embedding information E T Specifically:

[0074] 5-2-2-1 Initialization, iteration number m = 0, node embedding information

[0075] 5-2-2-2 Use activation function to obtain node embedding information E T , the formula is as follows:

[0076]

[0077] Where Relu represents the activation function, represents the weight matrix, Perform symmetric normalization processing, Indicates the update of the feature matrix X after the mth iteration;

[0078] 5-2-2-3 Judgment and Are they the same? If so, output the final node embedding information. If not, update m=m+1 and return to step 5-2-2-2;

[0079] 5-2-3 In the network topology, a gate mechanism is used to determine whether the attention value between the feature vectors of each node extracted in the attention module is valid. An attValue of 1 indicates valid, while a value of 0 indicates invalid. This prevents the backward transmission of attention values corresponding to non-adjacent nodes or weakly related nodes in the network topology structure.

[0080]

[0081] in is the network topology G T The embedding vector of the middle node u1, u2, att() represents the attention mechanism, and θ represents the threshold;

[0082] 5-3 uses random attention mechanism to combine network topology space and node feature space to learn long-distance dependencies. It extracts the common information of network topology space and node feature space through GCN with parameter sharing to obtain the common embedding E FT The specific process is as follows:

[0083] 5-3-1 respectively for the adjacency matrix A F and A T Extract common embedding information through shared parameter GCN;

[0084] 5-3-2 Initialization, iteration number m = 0, public embedding information

[0085] 5-3-3 Use activation function to obtain public embedding information E FT , the formula is as follows:

[0086]

[0087]

[0088] Where Relu represents the activation function, is the shared weight matrix of the network topology space and the node feature space, Perform symmetric normalization processing, Respectively represent the updated feature matrix X after the mth iteration of the node feature space and the network topology space;

[0089] 5-3-4 Determine whether If so, output the final node embedding information If not, update m=m+1 and return to step 5-2-3;

[0090] 5-3-5 The sum of the two is averaged to get the joint embedding E of the node feature space and the network topology space FT ;

[0091] 5-4 Embed the node output in step 5-1 into information E F , the node embedding information E output in step 5-2 T And the public embedding E output from step 5-3 FT , and weighted to get the final embedding E; the formula is as follows:

[0092] E=W F E F +W T E T +W FT E FT (9)

[0093] Where W F 、W T 、W FT They are E F 、E T 、E FT The weight value of

[0094] Step (6): Use the meta-path module to mine the high-level semantic information in the meta-path and further update the network topology structure;

[0095] like Figure 6 As shown in Figure 2, the meta-path module combines the meta-path with the network topology and adopts the attention mechanism. Like the external feature module, it uses the gate mechanism in the message propagation to eliminate the influence of noise in the original graph on the topology structure. It aggregates multiple meta-path information to obtain the network topology based on the meta-path, which is represented by the relationship matrix X. M .

[0096] The specific process is as follows:

[0097] 6-1 Processing of single meta-path P information;

[0098] By learning the target node v of the commodity relationship graph G based on the meta-path neighborhood of P Complex semantic information; specifically:

[0099] Assume that P(u, v) represents the target node v and the v-element path neighborhood The meta-path of neighbor node u in The attention mechanism is used to perform weighted summation on the instances of the meta-path P of the target node, and the process is:

[0100] First, use the attention mechanism and the SoftMax function to obtain a normalized attention weight value for the meta-path instance. The process is to first use the attention mechanism and the SoftMax function to obtain a normalized attention weight value for the meta-path instance. The calculation formula is as follows:

[0101]

[0102] in is the transpose of the parameter vector of the metapath P, h v is the embedding vector of node v, is the node embedding of the neighbor node u in the neighborhood of v-element path, || is the vector concatenation operator, represents the normalized impact factor;

[0103] Based on Perform weighted summation on all meta-path instances, and finally output the single meta-path processing result through the activation function The calculation formula is as follows:

[0104]

[0105] 6-2 Aggregate multiple different meta-paths;

[0106] Metapath The meta-path of all nodes of type A The node vectors under are nonlinearly transformed and then averaged. The calculation formula is as follows:

[0107]

[0108] where Num A is the number of meta-paths with node type A; Tanh is the activation function, is the processing result of a single meta-path, b is the offset, |VA | is the number of nodes of type A;

[0109] Then, the attention mechanism is used to perform weighted summation on the vectors of all meta-paths of the target type node. The calculation formula is as follows:

[0110]

[0111] in is the parameterized attention vector of type A node;

[0112] 6-3 Update the feature matrix X according to the meta-path to obtain the relationship matrix X representing the new network topology M , the calculation formula is as follows:

[0113]

[0114] in It is the result of aggregating all types of meta-paths;

[0115] Step (7): Construct a graph convolution module based on the features X between nodes of the same type N , Features E between different types of nodes, relationship matrix X M , output classification results;

[0116] Step (8): Recommend products based on the product node classification results;

[0117] The edges of the products in the product relationship graph G classified by the graph convolution module are accumulated, and the user’s interest in the products in this category is judged based on the accumulated value, and then product recommendations are made.

Claims

1. A product recommendation method based on a heterogeneous graph neural network with parallel feature extraction, characterized by The method comprises the following steps: Step (1): Obtain e-commerce product data, including product attributes and user behavior; Step (2): Preprocess the e-commerce product data; Specifically: A product relationship graph G is constructed based on e-commerce product data, wherein products are nodes, product attributes are node attributes, and edges are user behaviors. The product relationship graph G is represented by a feature matrix X. An element value of 0 in the feature matrix X indicates that there is no direct connection between two product nodes, an element value greater than 0 indicates a direct connection between the two product nodes, and an element value of 1 corresponds to a user behavior that the user clicked on the two products within a period of time, an element value of 2 corresponds to a user behavior that the user added the two products to the shopping cart at the same time, and an element value of 3 corresponds to a user behavior that the user ordered the two products at the same time. Based on user behavior, extract meta-paths for each node in the product relationship graph G; Step (3): Initial type classification of feature matrix X: According to the original commodity classification in e-commerce, the same type of nodes and the original edge relationships of these nodes in the feature matrix X are extracted from the feature matrix X to form the same type feature matrix X1; Step (4): Use the internal feature module to extract the same type of feature matrix ; Step (5): Using the external feature module, the random attention mechanism is combined with the network topology space and the node feature space to learn the long-distance dependency relationship and extract the features between different types of nodes. ; Step (6): Use the meta-path module to mine the high-level semantic information in the meta-path, further update the network topology, and extract the relationship matrix ; Step (7): Use the graph convolution module to calculate the feature matrix of the same type , characteristics between different types of nodes , relationship matrix , output classification results; Step (8): Recommend products based on the classification results of product nodes; specifically, accumulate the edges of the same category of products classified by the graph convolution module in the product relationship graph G, judge the user's interest in the category of products based on the accumulated value, and then recommend products; Step (4) is specifically: 4-1 The internal feature module learns the features of nodes of the same type through the attention mechanism and updates the node features: Given a pair of nodes of the same type , , attention mechanism learning node , The correlation between the two , that is, node For Node contribution; The calculation formula is as follows: ； in , Node , The embedding vector of represents the attention mechanism; 4-2 pairs Normalize the nodes using the SoftMax function to get the weight of the nodes. , the formula is as follows: ； in Indicates the number of nodes; () represents the normalization function; 4-3 weight values of nodes Compare with the threshold, that is, judge the correlation between two nodes of the same type. If If the correlation is greater than the threshold, the two nodes of the same type are considered to be strongly correlated and no processing is performed. Otherwise, the two nodes of the same type are considered to be weakly correlated and the corresponding elements in the feature matrix X1 of the same type are updated to 0. 4-4 Output the updated feature matrix of the same type through a fully connected layer ; Step (5) is specifically: 5-1 uses the random attention mechanism combined with the node feature space to learn long-distance dependencies. The specific process is as follows: 5-1-1 In the node feature space, after node feature mapping, according to the node feature matrix X and node similarity , take the k nodes with the highest similarity to build a graph , whose adjacency matrix is ; The node similarity The cosine similarity is calculated as follows: ； in and Represents the adjacency matrix midpoint and nodes The eigenvector of 5-1-2 Feature matrix X based on adjacency matrix Continuously update to get node embedding information ; 5-2 Using random attention mechanism combined with network topology space to learn long-distance dependencies The specific process is as follows: 5-2-1 In the network topology space, establish the network topology structure G based on the characteristic matrix X T , whose adjacency matrix is ; 5-2-2 For the feature matrix X, according to the adjacency matrix Continuously update to get node embedding information ; 5-2-3 In the network topology, a gate mechanism is used to determine whether the attention value between the feature vectors of each node extracted in the attention module is valid. 1 means valid, 0 means invalid, thereby blocking the backward transmission of attention values corresponding to non-adjacent nodes or weakly related nodes in the network topology; ； in , is the network topology G T midpoint , The embedding vector of represents the attention mechanism, Indicates the threshold value; 5-3 uses random attention mechanism to combine network topology space and node feature space to learn long-distance dependencies. It extracts the common information of network topology space and node feature space through GCN with parameter sharing to obtain common embedding The specific process is as follows: 5-3-1 respectively for the adjacency matrix and Extracting common embedding information through shared parameter GCN; 5-3-2 Initialization, iteration number m=0, public embedding information = feature matrix X; 5-3-3 Using activation functions to obtain public embedding information , the formula is as follows: ； ； Where Relu represents the activation function, is the shared weight matrix of the network topology space and the node feature space, Perform symmetric normalization processing, Respectively represent the updated feature matrix X after the mth iteration of the node feature space and the network topology space; 5-3-4 Determine whether = 、 = , if so, output the final node embedding information , if not, update m=m+1 and return to step 5-2-3; 5-3-5 Add and average to get the common embedding of node feature space and network topology space ; 5-4 Embed the node output in step 5-1 into the information , the node embedding information output in step 5-2 And the public embedding output from step 5-3 , weighted to get the final embedding ; The formula is as follows: ； in 、 、 They are 、 、 The weight value of .

2. The method according to claim 1, characterized in that Step 5-1-2 is as follows: 5-1-2-1 Initialization, iteration number m=0, node embedding information = Feature Matrix X 5-1-2-2 Use activation function to obtain node embedding information , the formula is as follows: ； in Relu represents the activation function, represents the weight matrix, Perform symmetric normalization processing, Indicates the update of the feature matrix X after the mth iteration; 5-1-2-3 Judgment and Are they the same? If so, output the final node embedding information. , if not, update m=m+1 and return to step 5-1-2-2.

3. The method according to claim 1, characterized in that Step 5-2-2 is as follows: 5-2-2-1 Initialization, iteration number m=0, node embedding information = Feature Matrix X 5-2-2-2 Use activation function to obtain node embedding information , the formula is as follows: ； in Relu represents the activation function, represents the weight matrix, Perform symmetric normalization processing, Indicates the update of the feature matrix X after the mth iteration; 5-2-2-3 Judgment and Are they the same? If so, output the final node embedding information. , if not, update m=m+1 and return to step 5-2-2-2.

4. The method according to claim 1, characterized in that Step (6) is specifically: 6-1 Processing of single meta-path P information; Learn the target node of the commodity relationship graph G by learning the meta-path instance p of P Based on metapath neighborhood Complex semantic information; specifically: Assume that P( ) represents the target node as well as Metapath neighborhood Neighbor nodes within The meta path, The attention mechanism is used to perform weighted summation on the instances of the meta-path P of the target node, and the process is: First, use the attention mechanism and SoftMax function to obtain a normalized attention weight value for the meta-path instance. The calculation formula is as follows: ； in is the transpose of the parameter vector of the metapath P, is the embedding vector of node v, yes Neighbor nodes in the meta-path neighborhood Node embedding, is the vector concatenation operator, represents the normalized impact factor; Based on Perform weighted summation on all meta-path instances, and finally output the single meta-path processing result through the activation function , and its calculation formula is as follows: ； 6-2 Aggregate multiple different meta-paths; Metapath , the meta-path of all nodes of type A The node vectors under are nonlinearly transformed and then averaged. The calculation formula is as follows: ； in is the number of meta-paths with node type A; Tanh is the activation function, is the processing result of a single meta-path, b is the offset, is the number of nodes of node type A; Then, the attention mechanism is used to perform weighted summation on the vectors of all meta-paths of the target type node. The calculation formula is as follows: ； in is the parameterized attention vector of type A node; 6-3 Update the feature matrix X according to the meta-path to obtain the relationship matrix representing the new network topology , the calculation formula is as follows: ； in It is the result of aggregating all types of meta-paths.

5. A product recommendation system implementing the method according to any one of claims 1 to 4, characterized in that include: Data acquisition module, used to obtain e-commerce product data, including product attributes and user behavior; The data preprocessing module is used to preprocess the e-commerce commodity data to obtain the commodity relationship graph G, feature matrix X, and meta-path; The same type data extraction module is used to extract the same type of nodes and the original edge relationships of these nodes in the feature matrix X from the feature matrix X to form the same type feature matrix X1; Internal feature module, used to extract feature matrices of the same type ; The external feature module is used to use the random attention mechanism to combine the network topology space and the node feature space to learn long-distance dependencies and extract features between different types of nodes. ; Meta-path module, used to mine high-level semantic information in meta-paths, further update network topology, and extract relationship matrix ; Graph convolution module, used to generate graphs based on the same type of feature matrix , characteristics between different types of nodes , relationship matrix , output the classification results.

6. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 4.

7. A computing device comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method according to any one of claims 1 to 4 is implemented.

Images