CN117994007A - Social recommendation method based on multi-view fusion heterogeneous graph neural network - Google Patents

Abstract

The present invention discloses a social recommendation method based on a multi-view fusion heterogeneous graph neural network, which belongs to the technical field of social network recommendation, and includes the following steps: step 1, obtaining the relationship between users and commodities in the social network, and establishing a heterogeneous graph according to whether there is a link relationship between the user and the commodity; step 2, establishing an attribute topology decoupling module in the node view, inputting the established matrix into the attribute topology decoupling module, to the embedded representation of the user under the node view; step 3, establishing a node encoding module in the network mode view, and obtaining the embedded representation of the user under the network mode view; step 4, establishing a semantic fusion module in the semantic view, and obtaining the embedded representation of the user under the semantic view; step 5, performing a fusion strategy of multi-view node representation, to the vector representation of user information, and performing personalized recommendation. The present invention explores multi-granularity information in social networks from three perspectives to recommend more accurate commodities to users.

Description

A social recommendation method based on multi-view fusion heterogeneous graph neural network

Technical Field

The present invention belongs to the technical field of social network recommendation, and in particular relates to a social recommendation method based on multi-view fusion heterogeneous graph neural network.

Background technique

Social networks usually present complex topological structures, including the connection mode between nodes, the degree distribution of nodes, and community structure. These structures may be sparse, dense, or highly clustered and small-world. In heterogeneous graph neural networks, nodes can contain different types of attribute information. For example, in social networks, user nodes may contain attributes such as age, gender, interests, and hobbies, and product nodes may contain attributes such as price, category, and sales. Through heterogeneous graph neural networks, these diverse attribute information can be effectively combined for recommendation tasks or other applications.

In the prior art, heterogeneous graph structure learning (HGSL) fuses the user's attribute graph, semantic graph, and original structure graph to generate a more comprehensive social network structure. ie-HGCN directly takes the graph data of the social network as input and performs multi-layer convolution to learn user node representations for specific tasks. SR-HGN is carefully designed to learn node representations in social information networks. It achieves hierarchical learning by aggregating information at the user and type levels, eliminating the need for prior knowledge in semantic path selection. These prior art methods often have unsatisfactory classification and clustering effects on users in social networks due to excessive coupling of fine-grained information and insufficient exploration of multi-granularity information.

Summary of the invention

In order to solve the above problems, the present invention proposes a social recommendation method based on multi-view fusion heterogeneous graph neural network. Firstly, an attribute-topology decoupling strategy is designed, which converts the user's attributes and its topology respectively, thereby protecting the user's fine-grained information and avoiding the transmission of multi-source information between adjacent nodes. At the same time, a new social network structure learning strategy is designed, which fuses subgraphs under different semantic relationships through graph-level attention, so as to not only identify the importance of different semantics, but also capture the interaction between users and products.

The technical solution of the present invention is as follows:

A social recommendation method based on multi-view fusion heterogeneous graph neural network includes the following steps:

Step 1: Obtain the relationship between users and products in the social network, and establish a heterogeneous graph based on whether there is a link relationship between users and products;

Step 2: Establish an attribute topology decoupling module in the node view, input the established matrix into the attribute topology decoupling module, perform feature and topology decoupling encoding, and obtain the node embedding representation under the node view;

Step 3: Establish a node encoding module in the network mode view, perform node encoding, and obtain a node embedding representation in the network mode view;

Step 4: Establish a semantic fusion module in the semantic view, perform semantic fusion, and obtain a node embedding representation under the semantic view;

Step 5: Implement a fusion strategy for multi-perspective node representations, use a splicing strategy to fuse information of different granularities in the social network, obtain a vector representation of user information, and perform personalized recommendations.

Furthermore, the specific process of step 1 is as follows: the heterogeneous graph matrix is established Indicates, /> , Represents the total number of user nodes, corresponding to the total number of rows in the matrix; /> Indicates the total number of commodity nodes, corresponding to the total number of columns of the matrix; the rows of the matrix represent user nodes, and the columns represent commodity nodes. If the first/> User nodes and /> If there is a link between the commodity nodes, then the matrix's Line No./> The column value is set to 1 if yes, otherwise 0.

Furthermore, the specific process of step 2 is as follows:

Step 2.1: The matrix Input attribute topology decoupling module, for a type of /> User node /> , through the mapping matrix/> Put the user node /> Attributes of /> Projected to a common dimension interval, as follows:

(1);

in, Is a user node/> The projection characteristics of is the dimension; /> is the activation function; /> represents vector offset;

Step 2.2: User node Projection characteristics of /> and topology/> The multi-layer perceptron is used to encode the common dimension interval, as follows:

(2);

(3);

in, is the encoded user node/> The attribute information of is the encoded user node/> The topological information representation of It is a multi-layer perceptron;

Step 2.3: Concatenate the two encoded representations, and then map the attribute information of users and products to the shared dimension through a multi-layer perceptron and activation function, and finally obtain the user node in the node view. Embedded representation of /> ,details as follows:

(4);

in, Represents a concatenation operation.

Furthermore, the specific process of step 3 is as follows:

Set user node With other types of nodes /> Connected, /> 、/> , 、/> Indicates different node types;

Apply node-level attention to Types of surrounding neighbors:

(5);

in, Is a user node/> The connection type is/> Neighbor node embedding representation; /> For user nodes/> The connection type is/> Neighbor nodes of; /> Is a neighbor node/> The projection characteristics of

For type/> Neighbor nodes/> For user nodes/> The attention is calculated as follows:

(6);

in, is an exponential function with the natural constant e as base;/> is the activation function; Is of type /> The node-level attention vector of is the transposition symbol; /> For user nodes/> Information; /> is the splicing symbol;

According to formula (5), we get the user node All different types of connected neighbor nodes are embedded ;

Using type-level attention, all different types of neighbor node embedding representations are fused together to obtain the user node under the network mode view. Embedded representation of /> ; The formula is as follows:

(7);

in, is the total number of types, /> The type index.

Furthermore, the specific process of step 4 is as follows:

Step 4.1: Divide the original social network into different graph structures through semantic knowledge, decompose the modeled graph data into different sub-modules to learn different semantic information; fuse different semantics through graph-level attention to form the integrated social network structure, which is calculated as follows:

(8);

in, It is the final relationship matrix obtained after merging different semantic relationships; /> Indicates that the parameter /> The graph-level attention layer is used to assign different weights to each exchange matrix; /> For semantic relations/> Matrix of; /> For semantic relations/> Matrix of; /> For semantic relations/> Matrix of

Step 4.2: Put it into the graph convolutional network GCN to obtain the node representation, and the calculation is as follows:

(9);

in, It is the user node in the semantic view/> Embedded representation of ; /> is the normalized relationship matrix; /> is a user feature; /> is the weight parameter.

Furthermore, the specific process of step 5 is as follows:

The embedded representations of the user nodes under the three views are spliced together, and the final embedded representation of the node is obtained through a multi-layer perceptron:

(10);

in, For user nodes/> The final embedding representation is the learned user information representation, and then personalized recommendations or predictions of user preferences are made based on the user's behavior and attribute information.

The beneficial technical effects brought by the present invention are as follows: by processing the user's features and topological information separately, the excessive coupling of fine-grained information caused by the message transmission of multi-source heterogeneous information can be decoupled, thereby realizing the separate modeling of user attribute information and topological information, decoupling the excessive coupling of the user's own fine-grained information, and applying this technology to the heterogeneous graph neural network model to improve the performance of node classification and clustering; three perspectives are designed to explore the multi-granular information in the social network, the user's own fine-grained information is extracted through the node perspective, the topological information of the user's neighbors is extracted through the network mode perspective, and the user's high-order heterogeneous information is captured through the semantic perspective, thereby realizing the exploration of multi-granular information in the social network. The present invention uses heterogeneous graph neural networks to simultaneously learn the simple topological structure and complex semantic relationship between users and commodities, and realizes the fusion of different information through splicing strategies, thereby improving the accuracy of user classification and clustering, and recommending more accurate commodities to users.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a social recommendation method based on a multi-view fusion heterogeneous graph neural network according to the present invention.

FIG. 2 is a schematic diagram showing the complexity comparison results of the model of the present invention and other models.

Detailed ways

The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

The present invention takes social networks as the research object, and improves the classification and clustering of users in social networks as the core goal. The key technical problems that need to be solved include: first, the attribute information of users is coupled. Second, the multi-granularity information of social networks is not fully mined. Solving these two problems can complete the modeling of complex social networks.

Therefore, the specific key issues to be solved by the present invention are:

Key technical issue 1: User attribute information is coupled;

The present invention designs the node perspective and the network mode perspective to decouple the entanglement between users and commodities caused by heterogeneity. Specifically, the node view models the user's attributes and decoupled information respectively to obtain the user's fine-grained information. The network mode view obtains the user's local topological information by modeling the commodity attributes and structural information of the user's surrounding area.

Key technical issue 2: Multi-granular information in social networks that is not fully mined;

The present invention designs three perspectives: node perspective, network pattern perspective and semantic perspective to learn different semantic information and mine different granular information of users in social networks. The node perspective decouples and learns the semantic structures of users and commodities respectively; the network pattern perspective learns the network pattern between users and commodities, and the semantic perspective learns the complex semantics of users and commodities. Through these three perspectives, the attribute information and topological structure information of users in social networks, as well as high-order heterogeneous structure information, are captured, realizing the mining of multi-granular information in social networks.

As shown in FIG1 , the present invention proposes a social recommendation method based on a heterogeneous graph neural network with multi-view fusion, which comprises the following steps:

Step 1: Obtain the relationship between users and products in the social network, and establish a heterogeneous graph based on whether there is a link relationship between users and products. The specific process is: a heterogeneous graph is a network that represents a complex relationship network. A heterogeneous graph is an abstract representation, and the network data needs to be modeled in the form of a matrix vector to transform the abstract into the concrete, so that knowledge can be applied to learning. Therefore, the heterogeneous graph established by the present invention is represented by a matrix. Indicates, /> ,/> Represents the total number of user nodes, corresponding to the total number of rows in the matrix; /> Indicates the total number of commodity nodes, corresponding to the total number of columns of the matrix. The rows of the matrix represent user nodes, and the columns represent commodity nodes. If the first/> User nodes and /> If there is a link between the commodity nodes, then the matrix's Line The column value is set to 1 if yes, otherwise 0.

Step 2: Establish an attribute topology decoupling module in the node view, input the established matrix into the attribute topology decoupling module, perform feature and topology decoupling encoding, and obtain the node embedding representation under the node view. The node view models user attributes and topology respectively to obtain fine-grained information. In order to avoid over-coupling of user attributes and topology, heterogeneous message passing is split into two modules for separate processing. In the attribute-topology decoupling module, only the user's attributes and adjacent topologies are transformed respectively. In this way, only the user's fine-grained information is retained, avoiding the transmission of multi-source heterogeneous information from adjacent structures. The specific process is:

Step 2.1: The matrix Input attribute topology decoupling module. Because there are different types of user nodes in social networks, the target features of user nodes are usually located in different spaces. Therefore, we first need to project all types of user node features into a common latent vector space. Specifically, for a type of/> User node /> , through a specific type of mapping matrix/> Put the user node /> Attributes of /> Projected to a common dimension interval, as follows:

(1);

in, Is a user node/> The projection characteristics of is the dimension; /> is the activation function; /> Stands for vector offset.

Attribute embedding can combine the attribute information and topological structure information of user nodes to obtain a more comprehensive user node representation, which can make user node embedding more accurate and reliable.

Step 2.2: User node Projection characteristics of /> and topology/> Encoded into a common dimension interval through a multi-layer perceptron (MLP), as follows:

(2);

(3);

in, is the encoded user node/> The attribute information of is the encoded user node/> The topological information representation of It is a multi-layer perceptron.

Step 2.3: Concatenate the two encoded representations, and then map the attribute information of users and products to the shared dimension through a multi-layer perceptron and activation function, and finally obtain the user node in the node view. Embedded representation of /> ,details as follows:

(4);

in, Represents a concatenation operation.

Step 3: Establish a node encoding module in the network model view, perform node encoding, and obtain the node embedding representation under the network model view. The network model view obtains the user's local information by modeling the attributes and structural information of the user's surrounding neighborhood. In the node encoding module of the network model, a two-level attention mechanism is used to aggregate the user's information. The specific process is:

The goal of the node encoding module is to learn the user node embedding in the network mode. Assuming that the user node With other types of nodes /> Connected, /> 、/> 、/> 、/> Indicates different node types, /> is the total number of types, /> is the type index, then the user node/> The connection type is/> The neighbor nodes of can be expressed as . For user nodes/> Different types of neighbors contribute differently to the representation of user nodes, and the contributions of neighbor nodes of the same type are also different. Therefore, node-level and type-level attention mechanisms are used to aggregate information from other types of neighbors to user nodes in a hierarchical manner. news.

Specifically, we first apply node-level attention to Types of surrounding neighbors:

(5);

in, Is a user node/> The connection type is/> Neighbor node embedding representation; /> For user nodes/> The connection type is/> Neighbor nodes of; /> Is a neighbor node/> The projection characteristics of

For type/> Neighbor nodes/> For user nodes/> The attention is calculated as follows:

(6);

in, is an exponential function with the natural constant e as base;/> is the activation function; Is of type /> The node-level attention vector of is the transposition symbol; /> For user nodes/> Information; /> A splicing symbol.

In practical applications, no aggregation from Instead of collecting all neighboring nodes' information, some surrounding neighbor information is randomly selected for aggregation in each epoch. An epoch represents the number of updates when all training data in learning are used once. Specifically, if the type is /> The number of surrounding neighbors exceeds a preset threshold/> , randomly select neighbor type as/> As/> , and not all neighbor nodes will be selected; if the number of neighbor nodes is less than the threshold /> , then the type will be repeatedly selected as/> The neighbor nodes of are collected until a threshold is reached. In this way, each node is guaranteed to aggregate the same amount of information from its neighbors and the diversity of node embeddings is increased in each epoch.

According to formula (5), we get the user node All different types of connected neighbor nodes are embedded ;

Use mean fusion to fuse them together to obtain the user nodes in the network model view The embedding representation of ; The formula is as follows:

(7);

This hierarchical attention mechanism can distinguish the contributions of different types of nodes and nodes of the same type, achieving fine-grained representation learning.

Step 4: Establish a semantic fusion module in the semantic view, perform semantic fusion, and obtain an embedded representation under the semantic view. The semantic view uses graph-level attention to merge different subgraphs into a single graph to understand the user's high-level heterogeneous information. In order to mine multi-granular information, semantic knowledge is used in the semantic fusion module to aggregate high-order information, and channel attention is used to fuse different semantic knowledge to achieve semantic interaction. The specific process is as follows:

Step 4.1: Divide the original social network into different graph structures through semantic knowledge, decompose the modeled graph data into different sub-modules to learn different semantic information; fuse them through graph-level attention to form the integrated social network structure, which is calculated as follows:

(8);

in, It is the final relationship matrix obtained after merging different semantic relationships; /> Indicates that the parameter /> The graph-level attention layer is used to assign different weights to each exchange matrix; /> For semantic relations/> Matrix of; /> For semantic relations/> Matrix of; /> For semantic relations/> The matrix of .

Step 4.2: Put it into the graph convolutional network GCN to get the node representation (in fact,/> It can be applied to any graph neural network GNN encoder), and is calculated as follows:

(9);

in, It is the user node in the semantic view/> Embedded representation of ; /> is the normalized relationship matrix; /> is a user feature; /> is the weight parameter.

Step 5: Implement a fusion strategy for multi-perspective node representations, use a splicing strategy to fuse information of different granularities in social networks, obtain a vector representation of user information, and perform personalized recommendations. The specific process is as follows:

The embedded representations of the user nodes under the three views are effectively spliced together, and the final embedded representation of the node is obtained through a multi-layer perceptron:

(10);

in, For user nodes/> The final embedding representation is the learned user information representation, and then personalized recommendations or predictions of user preferences are made based on the user's behavior and attribute information.

In order to prove the feasibility and superiority of the present invention, the following comparative experiments were carried out.

Experiment 1: The above-mentioned attribute topology decoupling module, node encoding module, semantic fusion module, and fusion strategy of multi-view node representation essentially constitute the model MHGNN of the present invention. The experimental data sets of the present invention include four data sets of ACM, IMDB, DBLP, and YELP. Each data set adopts four segmentation rates of 20%, 40%, 60%, and 80%. First, the results of the model MHGNN of the present invention and the seven baseline models of HAN, GTN, MAGNN, HGSL, HPN, RoHe, and ie-HGCN in semi-supervised node classification were evaluated. Macro-F1 and Micro-F1 were selected as the evaluation index method. The specific data results are shown in Table 1. It is compared with seven representative algorithm models. Macro-F1 and Micro-F1 are two calculation methods of F1-score. F1-score is an indicator used to measure the accuracy of the binary classification model, which is used to measure the accuracy of unbalanced data. The higher the value of Macro-F1 or Micro-F1, the higher the accuracy of the model. Among them, HAN is divided into different semantic subgraphs by different semantic relationships between users and commodities, and then different subgraphs are fused with user representations under different subgraphs through attention. GTN iteratively convolves user information through the relationship matrix between users and products. MAGNN retains all node attribute information by encoding the relationship between users and products in the form of sequences, but during the encoding process, user information is repeatedly encoded multiple times, resulting in information redundancy. HGSL learns the correct link relationship by utilizing user attributes and neighborhood structure to eliminate the noise information in the original graph topology. HPN alleviates the semantic confusion problem after multi-layer convolution by retaining the initial relationship matrix at each layer. RoHe still works efficiently under noise attacks by introducing the form of attention edges. ie-HGCN iteratively couples the information of the surrounding neighborhood layer by layer through relational convolution. The present invention processes social network data into a suitable format and feeds it back to the model to test the superiority of our proposed method.

Table 1 Results of each model on semi-supervised node classification (%);

.

As shown in Table 1, the MHGNN model of the present invention performs well on the four datasets of ACM, IMDB, DBLP, and YELP. In general, the user representation obtained by the MHGNN model of the present invention by aggregating information of different granularities from three perspectives is better than the representation obtained from a single perspective through semantic paths or modeling the user's own information. This is because the MHGNN model not only uses the semantic perspective to capture high-order heterogeneous information, but also uses the network model perspective and decoupling of features and topology to make the learned user representation have stronger representation capabilities.

Experiment 2: The present invention also conducts a node clustering experiment to verify the performance of the MHGNN model of the present invention. Specifically, the obtained user representation is put into K-means for clustering, and the number of clusters is set to the actual number of categories of the node. The results of two evaluation index methods, the average normalized mutual information (NMI) and the adjusted Rand index (ARI), are reported. The higher the value of NMI or ARI, the better the model performance.

Table 2 Results of MHGNN on node clustering (%);

.

As can be seen from Table 2, MHGNN has achieved good performance compared with other baseline models. Compared with other single-view models, MHGNN fully captures the semantic relationship between nodes and surrounding neighborhoods, making the distinction between users and neighborhood nodes higher and the clustering effect more obvious.

Figure 2 compares and analyzes the complexity of the proposed model MHGNN and six models including HAN, MAGNN, HGSL, HPN, RoHe, and ie-HGCN. As can be seen from Figure 2, although MHGNN uses three perspectives to learn the representation of users, it does not consume too much memory and time, because the three views are finally concatenated, and the subgraphs under different semantic paths are merged into a metagraph using attention, which saves the calculation time of each subgraph, reduces the amount of calculation, and improves the calculation efficiency.

Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims (6)

1. A social recommendation method based on multi-view fusion heterogeneous graph neural network, characterized by comprising the following steps: Step 1: Obtain the relationship between users and products in the social network, and establish a heterogeneous graph based on whether there is a link relationship between users and products; Step 2: Establish an attribute topology decoupling module in the node view, input the established matrix into the attribute topology decoupling module, perform feature and topology decoupling encoding, and obtain the node embedding representation under the node view; Step 3: Establish a node encoding module in the network mode view, perform node encoding, and obtain a node embedding representation in the network mode view; Step 4: Establish a semantic fusion module in the semantic view, perform semantic fusion, and obtain a node embedding representation under the semantic view; Step 5: Implement a fusion strategy for multi-perspective node representations, use a splicing strategy to fuse information of different granularities in the social network, obtain a vector representation of user information, and perform personalized recommendations. 2. According to the social recommendation method based on multi-view fusion heterogeneous graph neural network in claim 1, the specific process of step 1 is: the heterogeneous graph matrix established Indicates, /> ,/> Represents the total number of user nodes, corresponding to the total number of rows in the matrix; /> Indicates the total number of commodity nodes, corresponding to the total number of columns of the matrix; the rows of the matrix represent user nodes, and the columns represent commodity nodes. If the first/> User nodes and /> If there is a link between the commodity nodes, then the matrix's Line The column value is set to 1 if yes, otherwise 0. 3. According to claim 2, the social recommendation method based on multi-view fusion heterogeneous graph neural network is characterized in that the specific process of step 2 is: Step 2.1: The matrix Input attribute topology decoupling module, for a type of /> User node /> , through the mapping matrix/> Put the user node /> Attributes of /> Projected to a common dimension interval, as follows: (1); in, Is a user node/> The projection characteristics of is the dimension; /> is the activation function; /> represents vector offset; Step 2.2: User node Projection characteristics of /> and topology/> The multi-layer perceptron is used to encode the common dimension interval, as follows: (2); (3); in, is the encoded user node/> The attribute information of is the encoded user node/> The topological information representation of It is a multi-layer perceptron; Step 2.3: Concatenate the two encoded representations, and then map the attribute information of users and products to the shared dimension through a multi-layer perceptron and activation function, and finally obtain the user node in the node view. Embedded representation of /> ,details as follows: (4); in, Represents a concatenation operation. 4. According to claim 3, the social recommendation method based on multi-view fusion heterogeneous graph neural network is characterized in that the specific process of step 3 is: Set user node With other types of nodes /> Connected, /> 、/> 、/> , Indicates different node types; Apply node-level attention to Types of surrounding neighbors: (5); in, Is a user node/> The connection type is/> Neighbor node embedding representation; /> For user nodes/> The connection type is/> Neighbor nodes of; /> Is a neighbor node/> The projection characteristics of For type/> Neighbor nodes/> For user nodes/> The attention is calculated as follows: (6); in, is an exponential function with the natural constant e as base;/> is the activation function; Is of type /> The node-level attention vector of is the transposition symbol; /> For user nodes/> Information; /> is the splicing symbol; According to formula (5), we get the user node All different types of connected neighbor nodes are embedded ; Using type-level attention, all different types of neighbor node embedding representations are fused together to obtain the user node under the network mode view. Embedded representation of /> ; The formula is as follows: (7); in, is the total number of types, /> The type index. 5. According to claim 4, the social recommendation method based on multi-view fusion heterogeneous graph neural network is characterized in that the specific process of step 4 is: Step 4.1: Divide the original social network into different graph structures through semantic knowledge, decompose the modeled graph data into different sub-modules to learn different semantic information; fuse different semantics through graph-level attention to form the integrated social network structure, which is calculated as follows: (8); in, It is the final relationship matrix obtained after merging different semantic relationships; /> Indicates that the parameter /> The graph-level attention layer is used to assign different weights to each exchange matrix; /> For semantic relations/> Matrix of; /> For semantic relations Matrix of; /> For semantic relations/> Matrix of Step 4.2: Put it into the graph convolutional network GCN to obtain the node representation, and the calculation is as follows: (9); in, It is the user node in the semantic view/> Embedded representation of ; /> is the normalized relationship matrix; /> is a user feature; /> is the weight parameter. 6. According to claim 5, the social recommendation method based on multi-view fusion heterogeneous graph neural network is characterized in that the specific process of step 5 is: The embedded representations of the user nodes under the three views are spliced together, and the final embedded representation of the node is obtained through a multi-layer perceptron: (10); in, For user nodes/> The final embedding representation is the learned user information representation, and then personalized recommendations or predictions of user preferences are made based on the user's behavior and attribute information.