High-order dependence modeling method based on heterogeneous graph convolutional network

Abstract

The invention discloses a high-order dependency modeling method based on a heterogeneous graph convolutional network. Firstly, data pre-training is carried out through a multi-layer perceptron, a high-order dependency relationship between nodes is captured according to simplex complex and time skewness random walk, a high-order adjacency matrix is generated, and the depth and breadth of information propagation are enhanced. Then, a homogeneity constraint matrix is introduced in the label propagation process, information transmission weights are adjusted according to homogeneity of nodes and edges, the influence among the same kind of nodes is enhanced, and interference among different kinds of nodes is restrained; and finally, multi-dimensional information is integrated through an adaptive fusion strategy, and the classification performance is further improved. By introducing the high-order dependency modeling and label propagation mechanism, the precision and robustness of node classification can be effectively improved, and the method is suitable for a complex heterogeneous information network.

Description

Technical Field

[0001] The present invention relates to the field of heterogeneous information technology, and in particular to a high-order dependency modeling method based on a heterogeneous graph convolutional network. Background Art

[0002] With the rapid development of information technology, graph data, as an important data structure for expressing entities and their relationships, has been widely used in many fields, including social networks, recommendation systems, knowledge graphs, information retrieval, etc. In these fields, node classification tasks are widely used in data mining and analysis. For example, in social networks, node classification can be used to identify user groups and interest divisions; in recommendation systems, node classification helps to accurately classify users and products, thereby improving the accuracy and personalization of recommendation systems. Node classification tasks usually rely on graph neural networks (GNNs) to learn the representation of nodes and then complete the classification task. GNNs can aggregate the information of neighbor nodes through the information transmission mechanism between nodes, thereby dynamically updating the representation of nodes, so that nodes can effectively learn the dependencies in the graph structure.

[0003] However, despite the success of GNN in many tasks, traditional GNN methods still have significant shortcomings when facing heterogeneous graph (HG) data in the real world. Many graph data in the real world are heterogeneous, that is, they contain multiple types of nodes and edges, while traditional GNNs usually assume that the graph is homogeneous and the nodes and edges are of the same type. Therefore, how to effectively process and model the complex semantics and relationships in heterogeneous information networks has become a key issue in the current research on graph neural networks. Heterogeneous Information Network (HIN) is a type of graph structure containing different types of nodes and edges. In a heterogeneous information network, nodes and edges not only have different types, but may also carry different attributes. For example, in a social network, nodes can represent different users, posts, comments, etc., while edges can represent friendships between users, interactions between users and posts, etc. This diversity makes traditional GNN methods difficult to apply directly, because they usually assume that nodes and edges are homogeneous using the simplex complex to model high-order interactions in the heterogeneous information network, that is, they do not consider the type information of nodes and edges and cannot effectively handle the complex relationships between node and edge types.

[0004] In addition, traditional graph neural networks, such as graph convolutional networks (GCNs) and GraphSAGE, mainly focus on how to use direct adjacency relationships between nodes to learn node representations, assuming that the dependencies between nodes are mainly expressed as adjacency relationships. These methods update the representation of nodes by aggregating the neighbor information of each node to the node. However, the nodes and edges in heterogeneous information networks are not only of various types, but also the dependencies between nodes may be more complex. The relationships between many nodes are not only direct adjacency relationships, but may also form complex interaction patterns through higher-order paths, common neighbors, or indirect dependencies. For example, in social networks, the relationship between user A and user B can be reflected not only through direct interactions, but also through indirect connections through common interests, topics, or friends, while in recommendation systems, the relationship between users and products may be indirectly connected through ratings, purchases, and other behaviors. These high-order dependencies and diverse interaction patterns make it impossible for traditional GNN methods to effectively capture the complex interactions and deep semantic relationships between nodes, affecting the performance of node classification tasks.

[0005] To address this problem, researchers have proposed methods such as the Heterogeneous Graph Convolutional Network (HGCN), which aims to design a graph neural network architecture that is more adaptable to heterogeneous information networks in response to the diversity and complexity of heterogeneous graphs. HGCN introduces node type embedding and edge type embedding, so that each type of node and edge can be represented and aggregated through different mechanisms, thereby effectively modeling the interaction between different types of nodes and edges in heterogeneous graphs. For example, HGCN can introduce an attention mechanism between different types of nodes and edges, so that the model can assign different weights to different types of neighbor nodes, thereby achieving more accurate information aggregation. In this way, HGCN can capture the diversity in the graph and the complex dependencies between nodes, improving the performance of node classification.

[0006] Although existing HGCN methods can capture high-order dependencies to a certain extent, they usually rely on the adjacency information and local structure of the graph, and it is difficult to fully mine the more complex high-order interaction information in the graph. In order to solve this problem, researchers have begun to focus on how to design more effective high-order dependency modeling methods, aiming to expand the perception range of the model so that the graph neural network can capture more levels of dependencies, thereby improving the performance of the model. Summary of the invention

[0007] In order to solve the limitations and defects of the prior art, the present invention provides a high-order dependency modeling method based on a heterogeneous graph convolutional network, comprising: Step S1, pre-training nodes in a heterogeneous information network according to a multi-layer perceptron to learn deep representations of the nodes, wherein the deep representations are used to capture relationship information and attribute information between the nodes, as well as initial label information; Step S2, using simplex complexes to model high-order interactions in the heterogeneous information network, and using time-skewed random walks to quantify the dependencies between the nodes and high-order simplexes; Step S3: Use the label propagation mechanism to construct a homogeneity constraint matrix, which is used to homogenize the heterogeneous graph structure and node attributes, and use the pre-trained label information for back propagation learning to obtain a propagation path matrix of class information, which is used to constrain features and topological structures; Step S4: dynamically adjust the degree of attention to different nodes and features according to the adaptive attention mechanism, learn and capture the high-order dependencies between the nodes, and process heterogeneous features at the same time; Step S5: Based on the high-order dependency and the label propagation mechanism, a graph convolutional network is used to represent the graph structure and node features.

[0008] Optionally, the step S1 includes: Use the multi-layer perceptron to pre-train data. The output of the layer is expressed as follows: ; in, It is The output of the layer, represents the activation function, is the weight matrix; Through the forward propagation of the multi-layer perceptron, the model can learn the feature representation of the node layer by layer; During the pre-training process, the model optimizes the weight of each layer according to the input features of the node to obtain the probability matrix of the node And the label information matrix Y, the expression is as follows: ; Among them, S is the probability matrix, is the final output embedding of the multi-layer perceptron, is the activation function used to convert the output of the neural network into a probability matrix, is the probability matrix The product of its transpose is used to analyze the correlation between features, and Y is the label information matrix.

[0009] Optionally, step S2 includes: Modeling high-order interactions in the heterogeneous information network using the simplicial complex, dividing into high-order partial graphs according to the order of the simplicial complex; capturing dependencies between the nodes and the higher-order simplex; The dependency relationship between the node and the high-order simplex is quantified using the time skewness random walk to obtain a quantized value of the dependency relationship between the node and the high-order simplex, and the expression is as follows: ; in, Indicated in Time Node and simplex The high-level interactions between represents the order of the simplex, express The set of simplexes of order, It is a slave node arrive Simplex The correlation matrix of It is a slave node To simplex The correlation matrix of Is to introduce timestamp constraints and degree deviation The subsequent weight coefficient is expressed as follows: ; Among them, times ( )represent Time information in the data set, times ( )represent The time information in the data set, the timestamp constraint is used to ensure that the information is only propagated to future nodes or high-order simplexes, and the degree deviation is used to adjust the parameters Control the weight of node degree on random walk; Get the high-order adjacency matrix The expression is as follows: ; in, is a high-order adjacency matrix, is the adjustment parameter of the degree deviation, Represents nodes and corresponding The incidence matrix of a simplex of order .

[0010] Optionally, step S3 includes: The label information and the high-order adjacency matrix are used to propagate class-aware information, and the expression is as follows: ; in, is the label information matrix, yes adjacency matrix, is a high-order adjacency matrix, is the inverse of the degree matrix, is a learnable homogeneity constraint matrix, which is used to filter information based on the homogeneity of nodes and edges by adjusting the information transfer weights between nodes; The homogeneity constraint matrix is ​​learned and optimized through the label propagation algorithm. The optimization goal is to minimize the loss function of label propagation, which is expressed as follows: ; in, is the loss function of the label propagation, represents the predicted label, represents the true label.

[0011] Optionally, step S4 includes: The weight of the high-order adjacency matrix is ​​adjusted using the adaptive attention mechanism, and heterogeneous information and high-order dependency information are adaptively fused to achieve the fusion of multi-dimensional information, where the multi-dimensional information includes pairwise interaction information, high-order interaction information, homogeneous information, and heterogeneous information. The expression is as follows: ; in, is the probability matrix obtained by the multilayer perceptron, , is a high-order adjacency matrix, , and is the weight coefficient obtained through attention learning; Update the high-order adjacency matrix after fusion, the expression is as follows: ; in, is the adjacency matrix of the k-order neighborhood.

[0012] Optionally, step S5 includes: According to the graph convolutional network, node classification is performed, and the expression of the node representation is as follows: ; in, Represents the graph convolutional network The output of the layer, and It is The weight of the layer, and is the weight coefficient; The optimization target of the node classification is obtained, and the expression is as follows: ; in, is the loss of the graph convolutional network, is the loss of the multilayer perceptron, is the label propagation loss, and is the regularization parameter.

[0013] The present invention has the following beneficial effects: The present invention provides a high-order dependency modeling method based on a heterogeneous graph convolutional network. First, data pre-training is performed through a multi-layer perceptron, and the high-order dependency relationship between nodes is captured according to the simplex complex and time skewness random walk, and a high-order adjacency matrix is ​​generated to enhance the depth and breadth of information propagation. Then, a homogeneity constraint matrix is ​​introduced in the label propagation process, and the information transfer weight is adjusted according to the homogeneity of nodes and edges to enhance the influence between nodes of the same type and suppress the interference between nodes of different types. Finally, multi-dimensional information is integrated through an adaptive fusion strategy to further improve the classification performance. By introducing high-order dependency modeling and label propagation mechanisms, the present invention can effectively improve the accuracy and robustness of node classification, and is suitable for complex heterogeneous information networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a flowchart of an algorithm provided in Embodiment 1 of the present invention.

[0015] Figure 2 A line chart comparing node classification results provided in Example 1 of the present invention.

[0016] Figure 3 This is a visualization comparison of the classification results of the high-order dependency modeling method, H2GCN and HOG-GCN on the Cora dataset provided in Example 1 of the present invention.

[0017] Figure 4 This is a visualization comparison of the classification results of the high-order dependency modeling method, H2GCN and HOG-GCN on the Texas dataset provided in Example 1 of the present invention.

[0018] Figure 5 This is a visualization diagram of the ablation experiment provided in Example 1 of the present invention on the homogeneous dataset Citeseer.

[0019] Figure 6 This is a visualization diagram of the ablation experiment provided in Example 1 of the present invention on the heterogeneous dataset Wisconsin. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the technical solution of the present invention, the high-order dependency modeling method based on heterogeneous graph convolutional network provided by the present invention is described in detail below with reference to the accompanying drawings.

[0021] Embodiment 1 This embodiment aims at the problem that there are multiple types of nodes and edges in the network, the relationship between nodes and the information propagation mechanism are relatively complex, and the traditional graph neural network method is difficult to deal with. A high-order dependency modeling method based on a heterogeneous graph convolutional network is proposed. First, data pre-training is performed by a multi-layer perceptron (MLP), and the high-order dependency relationship between nodes is captured by combining simplex complex and time skewness random walk, and a high-order adjacency matrix is ​​generated, thereby enhancing the depth and breadth of information propagation. Then, a homogeneity matrix is ​​introduced in the label propagation process, and the information transmission weight is adjusted according to the homogeneity of nodes and edges, so as to enhance the influence between nodes of the same type and suppress the interference between nodes of different types. Finally, multi-dimensional information is integrated through an adaptive fusion strategy to further improve the classification performance. Experimental results show that this embodiment can effectively improve the accuracy and robustness of node classification by introducing high-order dependency modeling and label propagation mechanism, and is suitable for complex heterogeneous information networks, such as social network analysis, knowledge graphs and other fields.

[0022] This embodiment provides a high-order dependency modeling method based on a heterogeneous graph convolutional network, aiming to overcome the shortcomings of the prior art, and improve the node classification performance of the heterogeneous graph convolutional network in a complex heterogeneous information network environment by introducing simplicial complex modeling of high-order dependencies, time-biased random walks and label propagation mechanisms.

[0023] This embodiment proposes a high-order dependency modeling method based on a heterogeneous graph convolutional network, which includes: a data pre-training process, high-order dependency relationship modeling based on time skewness random walk, label propagation, an adaptive fusion strategy and a node classification model.

[0024] Data pre-training: Before node classification, multi-layer MLP is used to pre-train nodes in heterogeneous information networks to learn deep representations of nodes. These representations can capture the complex relationships and attribute information between nodes, and can also obtain initial label information. Through pre-training, the model can have richer node representations at the beginning of training, which lays the foundation for subsequent high-order dependency modeling and label propagation. The pre-training process uses a multi-level feature learning mechanism to help capture graph structure features at different levels, thereby improving classification accuracy.

[0025] Modeling of high-order dependencies based on temporal skewness random walks: In the model building phase, simplex complexes are used to model high-order interactions in heterogeneous information networks. Then, temporal skewness random walks are used to quantify the dependencies between 0-simplexes (i.e., nodes) and higher-order simplexes (i.e., higher-order relationships containing multiple nodes). By introducing temporal skewness to the random walk process, the directionality of information propagation can be dynamically adjusted according to the importance of the nodes in the network. The introduction of temporal skewness allows the model to give priority to node relationships with more recent interactions or more structural importance when propagating information, thereby improving the breadth and depth of information propagation.

[0026] Label propagation mechanism: In the information propagation stage, the label propagation mechanism is used to construct the homogeneity constraint matrix, and the pre-trained label information is used for back propagation learning to obtain the propagation path matrix of the class information, which constrains the features and topological structure. The role of this constraint matrix is ​​to homogenize the heterogeneous graph structure and node attributes, reduce the noise from the interaction of irrelevant nodes, and thus improve the accuracy of node classification. The label propagation mechanism effectively enhances the propagation of similarity information between nodes, which helps to accurately predict the node category.

[0027] Adaptive fusion strategy: In order to further improve the modeling capabilities of high-order dependencies and heterogeneous features, an adaptive attention mechanism is introduced. This mechanism can dynamically adjust the degree of attention to different nodes and features, and learn and capture complex relationships between nodes. The adaptive attention mechanism enables the model to adaptively learn high-order dependencies between nodes and effectively process heterogeneous features, thereby improving the network's ability to represent complex node relationships.

[0028] Node classification model: Combining high-order dependency modeling and label propagation mechanism, a graph convolutional neural network is used to learn more accurate representations from graph structure and node features, thereby improving the robustness and accuracy of node classification.

[0029] In summary, this embodiment provides a high-order dependency modeling method based on heterogeneous graph convolutional networks to improve the accuracy and robustness of heterogeneous network node classification. In a heterogeneous network environment with dynamic changes and more noise, this method can better maintain performance stability and reduce the interference of irrelevant nodes on the classification results.

[0030] This embodiment provides a high-order dependency modeling method based on a heterogeneous graph convolutional network, which is used to improve the node classification performance of the heterogeneous graph convolutional network in a complex heterogeneous information network environment.

[0031] The specific implementation of this embodiment is as follows: (1) Use multi-layer MLP to pre-train the data. The output of the layer is as follows: ; Through the forward propagation of the multi-layer perceptron, the model can learn the high-level feature representation of the node layer by layer. During the pre-training process, the goal of the model is to optimize the weights of each layer according to the input features of the node to obtain a probability matrix that can better represent the node. And label information matrix Y: ; This embodiment provides a high-order dependency modeling method based on a heterogeneous graph convolutional network. The core of this method is to introduce high-order dependencies and label propagation mechanisms, aiming to significantly improve the node classification performance of heterogeneous graph convolutional networks in complex heterogeneous information network environments. By deeply mining the high-order dependencies between nodes and combining the label propagation mechanism, high-order information and heterogeneous information are adaptively learned. This embodiment can more accurately capture and utilize the features and structural information in the graph, thereby improving the accuracy and robustness of node classification, and is suitable for complex heterogeneous information network environments.

[0032] 2) Modeling high-order dependencies and calculating high-order adjacency matrices

[0033] After completing data pre-training, a simplex complex is used to model high-order interactions in a heterogeneous information network, and is divided into a high-order partial graph according to the order of the simplex complex. Through the simplex complex, this embodiment can capture the dependencies between 0-simplex (node) and higher-order simplex (such as 2-simplex, 3-simplex, etc.) between nodes, and quantify these dependencies using time skewness random walk. In this process, the formula for calculating the quantitative value describing the dependency between a node and a high-order simplex can be expressed as: ; in, Indicated in Time Node and simplex The high-level interactions between and Slave nodes To simplex , slave node To simplex The correlation matrix of Is to introduce timestamp constraints and degree deviation The weight coefficient after is calculated by the following formula: ; The timestamp constraint ensures that information propagates only to future nodes or higher-order simplexes. times( ) and times( ) represent and In order to ensure that information propagation complies with the time order, only when times ( )>times( ) can be used to spread information, thus ensuring that the information flow follows the time sequence and avoids time sequence contradictions. The degree deviation of the node is adjusted by The size of can control the influence of node degree on random walk weight, so that nodes with higher degrees have greater weight in the process of information propagation, enhancing the robustness of the model. Finally, after derivation, the high-order adjacency matrix can be obtained as: ; (3) Perform class information propagation and optimize the homogeneity constraint matrix T After data pre-training and high-order dependency modeling are completed, the obtained label information is combined with the high-order adjacency matrix to propagate class-aware information: ; in, yes adjacency matrix, is the inverse of the degree matrix, It is the transfer matrix of label propagation. It can filter information based on the homogeneity of nodes and edges by adjusting the information transfer weights between nodes. It is also called the homogeneity matrix. It needs to be obtained through optimization so that during the information propagation process, the influence of nodes of the same type is enhanced, while the interaction between nodes of different types is weakened, which helps to deal with complex node types and edge relationships in heterogeneous graphs. The optimization goal is to minimize the loss function of label propagation: ; (4) Adaptive fusion strategy In order to adaptively fuse heterogeneous information and high-order dependency information, the attention mechanism is used to adjust the weights of the adjacency matrix, thereby achieving the fusion of multi-dimensional information, including direct pairwise interactions, high-order interactions, and homogeneous and heterogeneous information.

[0034] ; The fused adjacency matrix is ​​updated by the following formula: ; This process ensures that different graph information is fully utilized and weightedly fused through adaptive weights.

[0035] (5) Node classification model In the node classification stage, combined with the above-mentioned high-order dependency modeling, label propagation and adaptive fusion, this embodiment uses a graph convolutional network (GCN) to classify nodes. The calculation formula for node representation is: ; Finally, the optimization goal of node classification is: ; in, is the loss of the graph convolutional network, is the loss of the multi-layer perceptron, is the label propagation loss, and is the regularization parameter. By comprehensively optimizing these loss functions, the final model can accurately classify nodes.

[0036] The high-order dependency modeling method based on the heterogeneous graph convolutional network proposed in this embodiment can improve the node classification performance of the heterogeneous graph convolutional network in a complex heterogeneous information network environment by introducing high-order dependency relationships, label propagation mechanisms and heterogeneous information fusion strategies, and has broad application prospects.

[0037] This embodiment proposes a high-order dependency modeling method based on a heterogeneous graph convolutional network, which aims to make up for the shortcomings of the existing HGCN method in modeling high-order interactive information by introducing simplicial complexes to represent high-order dependencies between nodes. This method captures high-order interactions between simplices through temporal skewness random walks, so that the model can not only capture direct adjacency relationships, but also effectively learn complex dependencies between high-order structures, thereby improving the accuracy and robustness of node classification. In addition, this embodiment also combines a label propagation mechanism to further mine the complex semantic information of heterogeneous information networks, so that the model can dynamically adjust the information propagation method when processing heterogeneous nodes, thereby more accurately capturing the complex semantics and high-order dependencies in the graph.

[0038] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A high-order dependency modeling method based on heterogeneous graph convolutional networks, characterized in that: include: Step S1, pre-training nodes in a heterogeneous information network according to a multi-layer perceptron to learn deep representations of the nodes, wherein the deep representations are used to capture relationship information and attribute information between the nodes, as well as initial label information; Step S2, using simplex complexes to model high-order interactions in the heterogeneous information network, and using time-skewed random walks to quantify the dependencies between the nodes and high-order simplexes; Step S3: Use the label propagation mechanism to construct a homogeneity constraint matrix, which is used to homogenize the heterogeneous graph structure and node attributes, and use the pre-trained label information for back propagation learning to obtain a propagation path matrix of class information, which is used to constrain features and topological structures; Step S4: dynamically adjust the degree of attention to different nodes and features according to the adaptive attention mechanism, learn and capture the high-order dependencies between the nodes, and process heterogeneous features at the same time; Step S5: Based on the high-order dependency and the label propagation mechanism, a graph convolutional network is used to represent the graph structure and node features.

2. The high-order dependency modeling method based on heterogeneous graph convolutional network according to claim 1 is characterized in that: The step S1 comprises: Use the multi-layer perceptron to pre-train data. The output of the layer is expressed as follows: ； in, It is The output of the layer, represents the activation function, is the weight matrix; Through the forward propagation of the multi-layer perceptron, the model can learn the feature representation of the node layer by layer; During the pre-training process, the model optimizes the weight of each layer according to the input features of the node to obtain the probability matrix of the node And the label information matrix Y, the expression is as follows: ； Among them, S is the probability matrix, is the final output embedding of the multi-layer perceptron, is the activation function used to convert the output of the neural network into a probability matrix, is the probability matrix The product of its transpose is used to analyze the correlation between features, and Y is the label information matrix.

3. The high-order dependency modeling method based on heterogeneous graph convolutional network according to claim 2 is characterized in that: The step S2 comprises: Modeling high-order interactions in the heterogeneous information network using the simplicial complex, dividing into high-order partial graphs according to the order of the simplicial complex; capturing dependencies between the nodes and the higher-order simplex; The dependency relationship between the node and the high-order simplex is quantified using the time skewness random walk to obtain a quantized value of the dependency relationship between the node and the high-order simplex, and the expression is as follows: ； in, Indicated in Time Node and simplex The high-level interactions between represents the order of the simplex, express The set of simplexes of order , It is a slave node arrive Simplex The correlation matrix of is from a node to a simplex The correlation matrix of Is to introduce timestamp constraints and degree deviation The subsequent weight coefficient is expressed as follows: ； ； Among them, times ( )represent Time information in the data set, times ( )represent The time information in the data set, the timestamp constraint is used to ensure that the information is only propagated to future nodes or high-order simplexes, and the degree deviation is used to adjust the parameters Control the weight of node degree on random walk; Get a high-order adjacency matrix The expression is as follows: ； in, is a high-order adjacency matrix, is the adjustment parameter of the degree deviation, Represents nodes and corresponding The incidence matrix of a simplex of order .

4. The high-order dependency modeling method based on heterogeneous graph convolutional network according to claim 3 is characterized in that: The step S3 comprises: The label information and the high-order adjacency matrix are used to propagate class-aware information, and the expression is as follows: ； in, is the label information matrix, yes adjacency matrix, is a high-order adjacency matrix, is the inverse of the degree matrix, is a learnable homogeneity constraint matrix, which is used to filter information based on the homogeneity of nodes and edges by adjusting the information transfer weights between nodes; The homogeneity constraint matrix is ​​learned and optimized through the label propagation algorithm. The optimization goal is to minimize the loss function of label propagation, which is expressed as follows: ； in, is the loss function of the label propagation, represents the predicted label, represents the true label.

5. The high-order dependency modeling method based on heterogeneous graph convolutional network according to claim 4 is characterized in that: The step S4 comprises: The weight of the high-order adjacency matrix is ​​adjusted using the adaptive attention mechanism, and heterogeneous information and high-order dependency information are adaptively fused to achieve the fusion of multi-dimensional information, where the multi-dimensional information includes pairwise interaction information, high-order interaction information, homogeneous information, and heterogeneous information. The expression is as follows: ； in, is the probability matrix obtained by the multilayer perceptron, , is a high-order adjacency matrix, , and is the weight coefficient obtained through attention learning; Update the high-order adjacency matrix after fusion, the expression is as follows: ； in, is the adjacency matrix of the k-order neighborhood.

6. The high-order dependency modeling method based on heterogeneous graph convolutional network according to claim 5 is characterized in that: The step S5 comprises: According to the graph convolutional network, node classification is performed, and the expression of the node representation is as follows: ； in, Represents the graph convolutional network The output of the layer, and It is The weight of the layer, and is the weight coefficient; The optimization target of the node classification is obtained, and the expression is as follows: ； in, is the loss of the graph convolutional network, is the loss of the multilayer perceptron, is the label propagation loss, and is the regularization parameter.

Images