A method, system, device and storage medium for e-commerce recommendation based on multi-channel heterogeneous graphs

Through the multi-channel heterogeneous graph e-commerce recommendation method, the multi-channel convolution embedded attention module and semantic graph convolution module extracts feature information in the e-commerce heterogeneous graph, solving the problem of the failure to effectively utilize heterogeneous information in the existing technology. The problem of this achieves more accurate e-commerce recommendation predictions.

CN118799039BActive Publication Date: 2024-12-03FUJIAN UNIV OF TECH
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411274272.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-12
Publication Date
2024-12-03
Estimated Expiration
2044-09-12

AI Technical Summary

Technical Problem

The existing recommendation system based on graph neural network failed to effectively utilize the complex heterogeneous nodes and relationship information in e-commerce heterogeneous graphs, and could not fully utilize the various interactive information between users and users, users and products, resulting in recommendation prediction. Inadequate accuracy.

Method used

Using the multi-channel heterogeneous graph e-commerce recommendation method, by obtaining the data types and activity data between users and products in the e-commerce heterogeneous graph, the multi-channel convolution embedding attention module and semantic graph convolution module are used to extract feature information. , and information fusion is performed through feature fusion module, and finally the binary cross entropy loss function is trained by negative sampling to improve the accuracy of recommended predictions.

Benefits of technology

Effectively obtain rich heterogeneous information between users and products in the e-commerce heterogeneous map, improve the accuracy of e-commerce recommendations, and improve the performance of the recommendation system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118799039B_ABST
    Figure CN118799039B_ABST
Patent Text Reader

Abstract

The present invention discloses an e-commerce recommendation method, system, device and storage medium based on a multi-channel heterogeneous graph, and relates to the field of artificial intelligence deep learning technology. The method comprises: obtaining an e-commerce heterogeneous graph; the e-commerce heterogeneous graph comprises data types, activity data and corresponding attribute data; the data types comprise users and commodities; the e-commerce heterogeneous graph is input into a multi-channel heterogeneous graph e-commerce recommendation model for feature extraction, and heterogeneous information between users and commodities is obtained, which is used to represent the e-commerce recommendation result; the multi-channel heterogeneous graph e-commerce recommendation model comprises a feature fusion module and a multi-channel convolution embedding attention module and a semantic graph convolution module respectively connected to the feature fusion module; the multi-channel convolution embedding attention module is constructed based on graph convolution and dot product attention mechanisms; the semantic graph convolution module is constructed based on a KAN framework. The present invention can effectively obtain rich heterogeneous information between users and commodities in an e-commerce heterogeneous graph, thereby improving the accuracy of e-commerce recommendation prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence deep learning technology, and in particular to a method, system, device and storage medium for e-commerce recommendation based on multi-channel heterogeneous graphs. Background Art

[0002] In e-commerce, recommendation systems play an important role and have become an indispensable part. Recommendation systems can help users find content of interest from a large number of products and are widely used in all walks of life. As a filtering system, the goal of the recommendation system is to present personalized information to users in order to improve user experience and increase business profits. The earliest recommendation model used the collaborative filtering (CF) method to capture the recommendation effect by directly calculating the similarity between users. However, with the rapid development of graph neural networks (GNNs), recommendation system methods based on graph neural networks have gradually emerged. These methods take advantage of the advantages of graph neural networks in processing structured data and exploring structural information, and have become the latest research direction in recommendation systems.

[0003] Graph neural networks have become a powerful tool for modeling graph structured data in recommendation systems. With the rapid development of the Internet, more and more rich information can be obtained, which can help recommendation models improve their performance. In the network, there are multiple objects and various types of relationships between them, which constitute a heterogeneous graph. Heterogeneous graphs contain rich semantic and structural information, which can be used for flexible modeling. Since heterogeneous graphs have rich semantic information, more and more studies have begun to apply them to recommendation systems to process complex multivariate heterogeneous data. In e-commerce, various types of interaction data are generated between users and products, such as browsing, collection, and purchase. The heterogeneous graphs formed by these data contain a lot of useful information and can be used to provide high-quality recommendations. Therefore, how to make full use of the heterogeneity and diversity of heterogeneous graphs, deeply mine the implicit information in the network and apply it to the recommendation system is an urgent problem to be studied. However, the existing methods of building recommendation systems based on graph neural networks have failed to effectively utilize the complex and heterogeneous node and relationship information in the network. Their objects of action are still limited to a single type of homogeneous graph, and they cannot make full use of the various interaction information between users and users, and between users and products. Summary of the invention

[0004] The purpose of the present invention is to provide an e-commerce recommendation method, system, device and storage medium based on a multi-path heterogeneous graph, which can effectively obtain rich heterogeneous information between users and products in the e-commerce heterogeneous graph, thereby improving the accuracy of e-commerce recommendation prediction.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] An e-commerce recommendation method based on multi-path heterogeneous graphs, comprising:

[0007] Obtain an e-commerce heterogeneous graph; the e-commerce heterogeneous graph includes data types, activity data and corresponding attribute data; the data types include users and commodities; the activity data include clicks, inquiries and purchases;

[0008] The e-commerce heterogeneous graph is input into the multi-way heterogeneous graph e-commerce recommendation model for feature extraction to obtain heterogeneous information between users and products; the heterogeneous information is used to represent the e-commerce recommendation results; the multi-way heterogeneous graph e-commerce recommendation model includes a feature fusion module and a multi-way convolutional embedding attention module and a semantic graph convolution module respectively connected to the feature fusion module; the multi-way convolutional embedding attention module is constructed based on graph convolution and dot product attention mechanisms; the semantic graph convolution module is constructed based on the KAN framework.

[0009] Optionally, the attribute data of the user includes nationality, gender and search log; the attribute data of the commodity includes category, price, number of searches, number of visits and purchase log.

[0010] Optionally, the e-commerce heterogeneous graph uses the data types as graph nodes, and uses the user's activities on the products as edges, relationships or links between nodes, so that the characteristics of the users and the products are distributed on the graph nodes.

[0011] Optionally, the feature extraction process is:

[0012] Inputting the e-commerce heterogeneous graph into the multi-way convolution embedding attention module, extracting the user's attention features for potential products, and obtaining a first extracted feature;

[0013] Inputting the e-commerce heterogeneous graph into the semantic graph convolution module, extracting semantic information, and obtaining a second extracted feature;

[0014] The first extracted features and the second extracted features are input into the feature fusion module, and feature fusion is performed through a splicing fusion mechanism to obtain heterogeneous information between users and products.

[0015] Optionally, before applying the multi-channel heterogeneous graph e-commerce recommendation model, it also includes: training the multi-channel heterogeneous graph e-commerce recommendation model by minimizing the binary cross entropy loss function through negative sampling as the objective function.

[0016] The present invention also provides an e-commerce recommendation system based on multi-path heterogeneous graphs, comprising:

[0017] A data collection unit is used to obtain an e-commerce heterogeneous graph; the e-commerce heterogeneous graph includes data types, activity data and corresponding attribute data; the data types include users and commodities; the activity data include clicks, inquiries and purchases;

[0018] The e-commerce recommendation unit is used to input the e-commerce heterogeneous graph into the multi-channel heterogeneous graph e-commerce recommendation model for feature extraction to obtain heterogeneous information between users and commodities; the heterogeneous information is used to represent the e-commerce recommendation results; the multi-channel heterogeneous graph e-commerce recommendation model includes a feature fusion module and a multi-channel convolution embedding attention module and a semantic graph convolution module respectively connected to the feature fusion module; the multi-channel convolution embedding attention module is constructed based on graph convolution and dot product attention mechanism; the semantic graph convolution module is constructed based on the KAN framework.

[0019] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the above-mentioned multi-path heterogeneous graph-based e-commerce recommendation method.

[0020] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned multi-path heterogeneous graph-based e-commerce recommendation method.

[0021] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0022] The present invention discloses an e-commerce recommendation method, system, device and storage medium based on a multi-channel heterogeneous graph, the method comprising obtaining an e-commerce heterogeneous graph; the e-commerce heterogeneous graph comprises data types, activity data and corresponding attribute data; the data types comprise users and commodities; the e-commerce heterogeneous graph is input into a multi-channel heterogeneous graph e-commerce recommendation model for feature extraction, and heterogeneous information between users and commodities is obtained, which is used to represent e-commerce recommendation results; the multi-channel heterogeneous graph e-commerce recommendation model comprises a feature fusion module and a multi-channel convolution embedding attention module and a semantic graph convolution module respectively connected to the feature fusion module; the multi-channel convolution embedding attention module is constructed based on graph convolution and dot product attention mechanisms; the semantic graph convolution module is constructed based on a KAN framework. The present invention can effectively obtain rich heterogeneous information between users and commodities in an e-commerce heterogeneous graph, thereby improving the accuracy of e-commerce recommendation prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0024] Figure 1 It is a schematic diagram of the process of the e-commerce recommendation method based on multi-channel heterogeneous graphs of the present invention;

[0025] Figure 2 Schematic diagram of algorithm logic of the e-commerce recommendation method based on multi-channel heterogeneous graphs in this embodiment;

[0026] Figure 3 Schematic diagram of algorithm operation of the multi-channel heterogeneous graph e-commerce recommendation method in this embodiment. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0028] The purpose of the present invention is to provide an e-commerce recommendation method, system, device and storage medium based on a multi-path heterogeneous graph, which can effectively obtain rich heterogeneous information between users and products in the e-commerce heterogeneous graph, thereby improving the accuracy of e-commerce recommendation prediction.

[0029] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] like Figure 1 As shown, the present invention provides an e-commerce recommendation method based on a multi-path heterogeneous graph, comprising:

[0031] Step 100: Obtain an e-commerce heterogeneous graph; the e-commerce heterogeneous graph includes data types, activity data and corresponding attribute data; the data types include users and commodities; the activity data include clicks, inquiries and purchases.

[0032] Step 200: Input the e-commerce heterogeneous graph into the multi-way heterogeneous graph e-commerce recommendation model for feature extraction to obtain heterogeneous information between users and products; the heterogeneous information is used to represent the e-commerce recommendation results; the multi-way heterogeneous graph e-commerce recommendation model includes a feature fusion module and a multi-way convolution embedding attention module and a semantic graph convolution module respectively connected to the feature fusion module; the multi-way convolution embedding attention module is constructed based on graph convolution and dot product attention mechanisms; the semantic graph convolution module is constructed based on the KAN framework.

[0033] Based on the above scheme, provide Figure 2-Figure 3 The detailed processing process is shown.

[0034] S1: This embodiment selects the e-commerce dataset Alibaba in the field of recommendation system analysis, randomly extracts 50% of the edges as the training set, and the remaining edges as the test set, and uniformly standardizes them. The Alibaba dataset consists of user and item behavior logs collected by the e-commerce platform (Alibaba.com). It contains two types of nodes (users and products) and three types of activities (clicks, queries, and purchases). Users have attributes such as nationality, gender, and search logs, and products have attributes such as category, price, number of searches, number of visits, and purchase logs. The users and products in the dataset can be regarded as nodes of the graph, and the user's activities on the products can be regarded as edges, relationships, or links between nodes. The user and product features are distributed on the nodes of the graph. In this way, the recommendation prediction problem is transformed into a link prediction problem for a heterogeneous graph. The node feature matrices of the e-commerce heterogeneous graph are extracted from the dataset respectively. , and the adjacency matrix composed of different types of edges , and ,in N is the number of nodes, d is the dimension of the node feature.

[0035] S2: The multi-channel heterogeneous graph e-commerce recommendation model mainly includes: multi-channel convolutional embedding attention module, semantic graph convolution module and feature fusion module.

[0036] (1) Implementation of multi-way convolutional embedding attention module:

[0037] The multi-way convolution embedding attention module is mainly composed of simplified multi-way convolution and dot product attention. First, a simplified convolution is constructed: Unlike traditional graph convolution, the nonlinear activation and matrix normalization of graph convolution are removed to reduce the amount of parameter transfer, thereby speeding up the learning rate of convolution. The simplified multi-way convolution is used to embed the multi-way relationships (edges) between users and products in the e-commerce heterogeneous graph:

[0038] (1)

[0039] (2)

[0040] (3)

[0041] in, , and It is the adjacency matrix composed of the three relationships of click, query and purchase in the e-commerce heterogeneous graph. , and is the learnable weight matrix corresponding to different relations, ,Z query ∈ N×dout and To learn the output feature representation of the three relationships of click, query and purchase, dout is the embedding dimension of the output.

[0042] Afterwards, the multi-way embeddings learned by graph convolution on the multi-way relations of the e-commerce heterogeneous graph are fused using dot product attention and normalized using softmax(·). In particular, dot product attention does not need to interact with the embedding and the adjacency matrix, so it can learn the global feature importance of nodes in the e-commerce heterogeneous graph. Learning the global feature importance of nodes in the e-commerce heterogeneous graph helps capture the attention information of users on all potential products on the e-commerce platform. The formula is as follows:

[0043] (4)

[0044] Where T is the matrix transpose operation, C is the number of relationship types in the e-commerce heterogeneous graph. In the Alibaba dataset, there are three relationships: click, query, and purchase. C is 3.

[0045] In addition, the dot product attention can be extended to a multi-head attention mechanism. Specifically, multiple normalized dot product attentions are performed independently, and then the feature representations of their outputs are averaged as the final representation of the output of the multi-way convolutional embedding attention module. The calculation process is:

[0046] (5)

[0047] in, is the node representation learned through the multi-head attention mechanism, head is the number of heads of the multi-head attention mechanism, W head ∈ dout×dout is the learnable parameter matrix.

[0048] (2) Implementation of semantic graph convolution module:

[0049] Due to the heterogeneity of the link relationship between users and products in the e-commerce heterogeneous graph, different types of relationships have adjacency matrices with different feature spaces. Therefore, this embodiment uses Kolmogorov-Arnold Networks (KAN) to encode and learn the heterogeneous information of the graph, aggregate the adjacency matrices composed of different types of relationships, and generate a new adjacency matrix A ∈ N×NThe learned adjacency matrix covers the connections or relationships between all node types. The process is as follows:

[0050] (6)

[0051] in, a , b and c are learnable parameters and use softmax (·) function for normalization, and Φ(·) is a learnable residual activation function.

[0052] In particular, unlike the traditional method of fixing an activation function, the activation function Φ(·) here can be learned. x As an input example, Φ( x ) can be expressed as:

[0053] (7)

[0054] in, w is a learnable parameter factor to better control the overall size of the activation function, splin e( x ) is a learnable univariate function that can be parameterized as a spline curve. When initializing, each activation function is initialized as a spline curve. splin e( x )≈0.

[0055] After that, graph convolution is used to learn the feature information between nodes and their neighbor nodes in the e-commerce heterogeneous graph, which can capture useful information about the interaction between users and the nearest related products on the e-commerce platform. The formula is as follows:

[0056] (8)

[0057] in, ReLU (·) is a nonlinear activation function, and are the graph convolution learnable weight matrices of the first and second layers respectively, Z G ∈ N×dout It is the node feature representation learned by the semantic graph convolution module.

[0058] (3) Implementation of feature fusion module:

[0059] In order to better learn the feature information and heterogeneous information of the e-commerce heterogeneous graph, this embodiment designs a spliced ​​feature fusion module to fuse the node representations learned by the multi-way convolution embedding attention module and the semantic graph convolution module, thereby obtaining the final node representation of the multi-way heterogeneous graph e-commerce recommendation model. Z ∈ N×2dout The formula is:

[0060] (9)

[0061] in, α and β are all learnable parameters and are used softmax (·) function to normalize, Norm (·) is to perform L2 regularization on the output feature representation to stabilize the training results. ReLU (·) is a nonlinear activation function.

[0062] S3: For heterogeneous input graph data G =( V , E ),in V is a collection of nodes, E is a set of edges. After S1, the node feature matrix of the e-commerce heterogeneous graph is extracted. X ∈ N×d , and the adjacency matrix composed of different types of edges A click ∈ N×N , A query ∈ N×N and A buy ∈ N×N , input into the model constructed by S2, and finally embed the final node Z Make recommendation predictions. The model training optimizes the model parameters by minimizing the binary cross entropy loss function through negative sampling as the objective function:

[0063] (10)

[0064] in, Z u and Z i The nodes are u and nodes i represents, T represents matrix transpose, σ (·) is the activation function sigmoid,s Can be any vector similarity metric function (such as inner product), Ω is the set of user-item positive node pairs, Ω- is the set of negative node pairs sampled from all unobserved user-item node pairs. The goal is to use the loss function to increase the similarity between node representations in positive samples while reducing the similarity between node representations in negative samples.

[0065] During the entire model training process, by adjusting hyperparameters such as dropout, weight-decay, and learning rate, we finally find the hyperparameter combination that makes the model perform best on the test data set.

[0066] S4: The established model uses the area under the ROC curve (AUROC) and the area under the precision-recall curve (AUPRC) as evaluation indicators to evaluate the feature learning ability of the model on the e-commerce dataset Alibaba. The experimental results achieve a better performance.

[0067] Therefore, this embodiment has the following beneficial effects:

[0068] This embodiment proposes an e-commerce recommendation method based on a multi-way heterogeneous graph and constructs a corresponding model. This method can effectively capture the rich heterogeneous information in the e-commerce heterogeneous graph. The e-commerce recommendation system based on the multi-way heterogeneous graph consists of a multi-way convolution embedded attention module, a semantic graph convolution module and a feature fusion module. The multi-way convolution embedded attention module first learns the feature information contained in the different interactive activities between users and commodities in the e-commerce heterogeneous graph through simplified graph convolution. The simplified graph convolution removes the nonlinearity and matrix normalization of the graph convolution to reduce the amount of parameter transfer, thereby accelerating the learning rate of the graph convolution. Then, multiple output feature representations of the simplified graph convolution are used as embeddings to calculate its dot product attention to capture the attention features of users to all potential commodities in the e-commerce heterogeneous graph. The semantic graph convolution module first uses Kolmogorov-Arnold Networks (KAN) to learn multiple relationships between users and commodities in the e-commerce heterogeneous graph, and generates an adjacency matrix containing the relationships between all users and commodity types. Then, the graph convolution method is used to capture the semantic information of the e-commerce heterogeneous graph. The feature fusion module fuses the feature representations learned by the multi-way convolution embedding attention module and the semantic graph convolution module through a splicing fusion mechanism. This embodiment can effectively obtain rich heterogeneous information between users and products in the e-commerce heterogeneous graph, thereby improving the accuracy of e-commerce recommendation prediction.

Claims

1. A method for e-commerce recommendation based on multi-path heterogeneous graphs, characterized in that: include: Obtain an e-commerce heterogeneous graph; the e-commerce heterogeneous graph includes data types, activity data and corresponding attribute data; the data types include users and commodities; the activity data include clicks, inquiries and purchases; The e-commerce heterogeneous graph is input into a multi-path heterogeneous graph e-commerce recommendation model for feature extraction to obtain heterogeneous information between users and products; The heterogeneous information is used to represent the e-commerce recommendation results; The multi-channel heterogeneous graph e-commerce recommendation model includes a feature fusion module and a multi-channel convolution embedding attention module and a semantic graph convolution module respectively connected to the feature fusion module; the multi-channel convolution embedding attention module is constructed based on graph convolution and dot product attention mechanism; the semantic graph convolution module is constructed based on the KAN framework; wherein, the construction process of the multi-channel heterogeneous graph e-commerce recommendation model is: (1) Implementation of multi-way convolutional embedding attention module: The multi-way convolution embedding attention module consists of a simplified version of multi-way convolution and dot product attention. First, a simplified convolution is constructed: unlike the traditional graph convolution, the nonlinear activation and matrix normalization of the graph convolution are removed to reduce the amount of parameter transfer. The simplified multi-way convolution is used to embed the multi-way relationship between users and products in the e-commerce heterogeneous graph: (1) (2) (3) in, , and It is the adjacency matrix composed of the three relationships of click, query and purchase in the e-commerce heterogeneous graph. , and is the learnable weight matrix corresponding to different relations, , Z query ∈ N×dout and To learn the output feature representation of the three relationships of click, query and purchase, dout is the embedding dimension of the output; Afterwards, the multi-way embeddings learned by graph convolution on the multi-way relations of the e-commerce heterogeneous graph are fused using dot product attention and normalized using softmax(·); dot product attention does not require the interaction between the embedding and the adjacency matrix, and the formula is as follows: (4) Where T is the matrix transpose operation, C is the number of relationship types in the e-commerce heterogeneous graph. In the Alibaba dataset, there are three relationships: click, query, and purchase. C is 3; In addition, the dot product attention is expanded to a multi-head attention mechanism, multiple normalized dot product attentions are performed independently, and then the feature representations of their outputs are averaged as the final representation of the output of the multi-way convolutional embedding attention module. The calculation process is: (5) in, is the node representation learned through the multi-head attention mechanism, head is the number of heads of the multi-head attention mechanism, W head ∈ dout×dout is a learnable parameter matrix; (2) Implementation of semantic graph convolution module: Due to the heterogeneity of the link relationship between users and products in the e-commerce heterogeneous graph, different types of relationships have adjacency matrices with different feature spaces. KAN is used to encode and learn the heterogeneous information of the graph, and the adjacency matrices composed of different types of relationships are brought together to generate a new adjacency matrix. A ∈ N×N , the process is as follows: (6) in, a , b and c are learnable parameters and use softmax (·) function is normalized, Φ(·) is the learnable residual activation function; by x As an input example, Φ( x ) is expressed as: (7) in, w is a learnable parameter factor to better control the overall size of the activation function, splin e( x ) is a learnable univariate function that can be parameterized as a spline curve. When initializing, each activation function is initialized as a spline curve. splin e( x )≈0; Afterwards, graph convolution is used to learn the feature information between nodes and their neighbor nodes in the e-commerce heterogeneous graph, capturing useful information about the interaction between users and the nearest related products on the e-commerce platform. The formula is as follows: (8) in, ReLU (·) is a nonlinear activation function, W (1) ∈ N×d and W (2) ∈ dout×dout are the graph convolution learnable weight matrices of the first and second layers respectively, Z G ∈ N×dout It is the node feature representation learned by the semantic graph convolution module; (3) Implementation of feature fusion module: The node representations learned by the multi-way convolution embedding attention module and the semantic graph convolution module are fused to obtain the final node representation of the multi-way heterogeneous graph e-commerce recommendation model. Z ∈ N×2dout , the formula is: (9) in, α and β are all learnable parameters and are used softmax (·) function to normalize, Norm (·) is to perform L2 regularization on the output feature representation to stabilize the training results. ReLU (·) is a nonlinear activation function.

2. The e-commerce recommendation method based on multi-channel heterogeneous graph according to claim 1 is characterized in that: The user's attribute data includes country, gender and search log; the commodity's attribute data includes category, price, search times, visit times and purchase log.

3. The e-commerce recommendation method based on multi-channel heterogeneous graph according to claim 1 is characterized in that: The e-commerce heterogeneous graph uses the data types as graph nodes and the user's activities on the products as edges, relationships or links between nodes, so that the characteristics of the users and the products are distributed on the graph nodes.

4. The e-commerce recommendation method based on multi-channel heterogeneous graph according to claim 1 is characterized in that: The processing process of feature extraction is: Inputting the e-commerce heterogeneous graph into the multi-way convolution embedding attention module, extracting the user's attention features for potential products, and obtaining a first extracted feature; Inputting the e-commerce heterogeneous graph into the semantic graph convolution module, extracting semantic information, and obtaining a second extracted feature; The first extracted features and the second extracted features are input into the feature fusion module, and feature fusion is performed through a splicing fusion mechanism to obtain heterogeneous information between users and products.

5. The e-commerce recommendation method based on multi-path heterogeneous graph according to claim 1 is characterized in that: Before applying the multi-channel heterogeneous graph e-commerce recommendation model, it also includes: training the multi-channel heterogeneous graph e-commerce recommendation model by minimizing the binary cross entropy loss function through negative sampling as the objective function.

6. An e-commerce recommendation system based on multi-path heterogeneous graphs, applied to the method according to any one of claims 1 to 5, characterized in that: include: A data collection unit is used to obtain an e-commerce heterogeneous graph; the e-commerce heterogeneous graph includes data types, activity data and corresponding attribute data; the data types include users and commodities; the related activity data include clicks, inquiries and purchases; An e-commerce recommendation unit, configured to input the e-commerce heterogeneous graph into a multi-path heterogeneous graph e-commerce recommendation model for feature extraction, and obtain heterogeneous information between users and commodities; The heterogeneous information is used to represent the e-commerce recommendation results; The multi-way heterogeneous graph e-commerce recommendation model includes a feature fusion module and a multi-way convolution embedding attention module and a semantic graph convolution module respectively connected to the feature fusion module; the multi-way convolution embedding attention module is constructed based on graph convolution and dot product attention mechanism; the semantic graph convolution module is constructed based on the KAN framework.

7. An electronic device, characterized in that: It includes a memory and a processor, the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the multi-path heterogeneous graph based e-commerce recommendation method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements the multi-channel heterogeneous graph e-commerce recommendation method as described in any one of claims 1 to 5.

Citation Information

Patent Citations

  • Shopping recommendation method and system based on heterogeneous graph neural network

    CN113191838A

  • Double-target cross-domain recommendation method and system based on knowledge graph and multi-head attention

    CN115760279A

  • Personalized commodity recommendation method and system based on multilayer heterogeneous graph convolution model

    CN115860880A

  • Personalized recommendation method and system based on heterogeneous graph attention and electronic equipment

    CN116680475A

  • Data processing method and device based on artificial intelligence, equipment and medium

    CN116701706A