Method and system for classifying citation network nodes of hybrid disparity map high-order neighbor convolutional network

Abstract

The invention discloses a citation network node classification method and a citation network node classification system for a hybrid disparity map high-order neighbor convolutional network. According to the method, an adjacent matrix of a graph is constructed through two main strategies: firstly, first-order to k-order neighbor information of nodes is fused through a mixed function, and a new adjacent matrix is adaptively learned; secondly, generating a soft label matrix of the nodes by utilizing a multi-layer perceptron model, and calculating a class similarity matrix among the nodes according to the soft label matrix; and in the high-order neighbor convolutional network model of the mixed disparity map, the two learned adjacency matrixes and class similarity matrixes are used for operation, so that more effective node feature updating is carried out. And then, optimization is carried out through a loss function, so that the accuracy of node classification during processing of the different illustration graph data is improved. And finally, through comparison experiments with other network models on a plurality of real data sets, the method can significantly improve the accuracy of node classification on the disparity map, and verifies the effectiveness of the hybrid disparity map high-order neighbor convolutional network in the aspects of node information extraction and node classification accuracy improvement.

Description

Technical Field

[0001] The present invention belongs to the field of graph convolutional neural networks, and in particular relates to a citation network node classification method and system for a mixed heterogeneous graph high-order neighbor convolutional network. Background Art

[0002] In the analysis of citation networks, graph neural networks (GNNs) have significantly promoted the development of node classification technology. In particular, GNNs perform well when dealing with citation network data with homogeneity, where documents tend to cite or be cited by other documents with similar topics or fields. However, when faced with citation networks with heterogeneity, that is, nodes (documents) in the network represent different disciplines or research fields, the performance of traditional GNNs is often limited. This is mainly because in the process of information aggregation, traditional models have difficulty in effectively distinguishing and utilizing the characteristics of documents from different disciplines.

[0003] When dealing with citation networks with complex connection patterns, the diverse association patterns between documents increase the challenge of classification tasks. For example, in some citation networks, there may be a large number of cross-disciplinary citations, which not only increases the difficulty of information aggregation, but also may affect the accuracy of the model in the classification process. In addition, existing graph neural network models usually assume that the connections in the network are highly homogeneous, which is obviously not applicable when dealing with citation networks with strong heterogeneity. A key assumption of traditional graph convolutional neural networks is that directly connected nodes in the network have similar features or categories, that is, the homogeneity assumption. In this framework, the feature update of a node is achieved by aggregating the information of its direct neighbors, aiming to enhance the feature representation of the target node through homogeneous neighbor feature propagation. However, when facing heterogeneous graphs, this assumption becomes no longer applicable. Nodes in heterogeneous graphs, even if they are interconnected, may exhibit completely different features and category labels. In this case, the key information often comes from the second-order or higher-order neighbors of the central node, and directly applying traditional graph convolutional networks may cause a lot of interference from irrelevant information.

[0004] Before the present invention, the processing of citation networks mainly relied on traditional graph neural network methods, such as simple graph convolutional networks and other basic graph processing techniques. These methods usually assume that the documents connected in the network tend to have similar research topics or belong to the same subject area. Therefore, the traditional method mainly updates the feature representation of the central node (document) by averaging or weighted averaging the features of neighboring nodes, which often leads to excessive smoothing of the document representation, thereby weakening the model's ability to capture differences between documents in processing citation networks with strong heterogeneity, limiting the application effect of graph neural networks in complex academic networks and other citation data with strong heterogeneity. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a citation network node classification method and system for a mixed heterogeneous graph high-order neighbor convolutional network to improve the accuracy of node classification.

[0006] Technical solution: To achieve the above-mentioned invention object, the present invention adopts the following technical solution:

[0007] A citation network node classification method for a hybrid heterogeneous graph high-order neighbor convolutional network comprises the following steps:

[0008] (1) Obtain a network graph dataset. The network graph is an undirected graph in which nodes have different attributes, which are represented by the node feature matrix X.

[0009] (2) Adaptively combine first-order to k-order neighbor information to learn a new adjacency matrix based on a mixing function

[0010] (3) Use the multi-layer perceptron MLP to generate the soft label matrix B of the nodes, and calculate the class similarity matrix S between the nodes based on it.

[0011] (4) Using the graph adjacency matrix of mixed multi-order neighbors Iteratively update the node features with the node class similarity matrix S to obtain the node embedding Where n is the number of nodes and c is the number of node categories.

[0012] (5) Embed the generated nodes Make node category prediction.

[0013] Furthermore, the step (2) calculates a new adjacency matrix This is achieved through the following formula:

[0014]

[0015] in is the relearned graph adjacency matrix, represents the learnable parameter matrix, Θ i T means Θ i The transpose of A i represents the i-order neighbor information matrix, when i=0, A 0 represents the identity matrix, k represents the longest neighbor class, i! refers to the factorial of i, and f represents the mixing function.

[0016] Furthermore, the workflow of calculating the node-to-node class similarity matrix in step (3) includes:

[0017] The node feature matrix X of the graph is input into a pre-trained multi-layer perceptron MLP and normalized with the softmax function to ensure that the sum of the probabilities of each node's soft label in all categories is 1 to generate the predicted category probability of the node and obtain the final soft label matrix Using the obtained soft label matrix B, we further calculate the class similarity matrix between nodes

[0018] Furthermore, the node embedding Z in step (4) is implemented by the following formula:

[0019]

[0020] Among them, Z (l) is the node feature representation of the lth layer, where Z (0) =X, and is the weight matrix of the l-1th layer, express The degree matrix of .

[0021] Furthermore, the step (5) generates node embedding in the model Finally, the softmax function is used to predict the node category.

[0022] Furthermore, by minimizing the loss function The parameters in the node classification method are obtained by training, where γ is a hyperparameter used to balance the two loss terms. is node v i The predicted label, y i ∈{0,1} 1×c is the true label, is node v i The characteristic vector of is the training set, and loss(·) is the cross entropy loss function.

[0023] The present invention also provides a citation network node classification system of a hybrid heterogeneous graph high-order neighbor convolutional network, comprising:

[0024] The data collection module is used to obtain the citation network graph dataset. The citation network graph is an undirected graph with different nodes, which is represented by the node feature matrix X.

[0025] Matrix update and convolution module, using a mixing function to adaptively combine first-order to k-order neighbor information to learn a new adjacency matrix Use the multi-layer perceptron MLP to generate the soft label matrix of the node and calculate the class similarity matrix S between the nodes; use the graph adjacency matrix of the mixed multi-order neighbors The node features are iteratively updated with the node class similarity matrix S to obtain the node embedding Where n is the number of nodes, c is the number of node classifications;

[0026] Prediction module, embedding the generated nodes Make node category prediction.

[0027] The present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the citation network node classification method of the mixed heterogeneous graph high-order neighbor convolutional network.

[0028] Beneficial effects: Compared with the prior art, the advantages and effects of the present invention are to make up for the deficiency that message transmission brings a large amount of noise data and affects the performance of downstream tasks, by introducing an adaptive new adjacency matrix and using the class similarity matrix obtained by the MLP model. The specific advantages are as follows:

[0029] (1) In heterogeneous graphs, directly connected nodes often do not belong to the same category. Nodes of the same category generally exist in the second or higher order. Traditional graph convolutional neural networks use the original adjacency matrix to update features by aggregating the information of first-order neighbors, which will introduce noise data and affect the accuracy of node classification. Therefore, the hybrid function f adaptive graph construction module path dynamically constructs and updates the adjacency matrix by considering the first-order to k-order neighbor information of the node and the potential relationship between the nodes; effectively integrates and updates the node features in the graph to improve the classification performance of the model.

[0030] (2) The node similarity calculation module is specifically used to evaluate the similarity between nodes in the graph, helping the model understand and process the complex node relationships in the graph, especially in the context of heterogeneous graphs. Traditional graph neural network models only focus on the feature information of the nodes themselves and ignore the rich connection relationships between nodes. The hybrid heterogeneous graph high-order neighbor convolutional network module realizes the effective integration of node feature information and the relationship between nodes, providing a comprehensive node analysis approach.

[0031] Aiming at the problem of noise data in traditional graph convolutional neural network methods, the present invention proposes a citation network node classification system of hybrid heterogeneous graph high-order neighbor convolutional network. The adjacency matrix of the graph is reconstructed through two strategies to adapt to the characteristics of heterogeneous graphs: one is to use the hybrid function f adaptive graphing module path to learn a new adjacency matrix by considering the first-order to k-order neighbor information of the node; the other is to use the MLP model to generate the soft label matrix of the node, and calculate the class similarity matrix between the nodes based on it. The two learned adjacency matrices and class similarity matrices are used to update the node features to better adapt to the characteristics of heterogeneous graphs. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the overall process of the method of the present invention.

[0033] Figure 2 Detailed structural diagram of the algorithm of the present invention. DETAILED DESCRIPTION

[0034] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0035] The main technical idea of ​​the present invention is: the citation network node classification method of the hybrid heterogeneous graph high-order neighbor convolutional network reconstructs the adjacency matrix of the graph through two main strategies: first, by fusing the first-order to k-order neighbor information of the node, a new adjacency matrix is ​​adaptively learned; second, the multi-layer perceptron model MLP is used to generate the soft label matrix of the node, and the class similarity matrix between the nodes is calculated based on it. In the citation network node classification system model of the hybrid heterogeneous graph high-order neighbor convolutional network, these two learned adjacency matrices and class similarity matrices are used to update node features more effectively; then a convolution operation is performed to obtain the node feature representation; finally, the node feature representation is predicted. Through comparative experiments with other network models on multiple real data sets, we have demonstrated the effectiveness of the citation network node classification system of the hybrid heterogeneous graph high-order neighbor convolutional network in extracting node information and improving the accuracy of node classification.

[0036] The embodiment of the present invention discloses a citation network node classification method of a hybrid heterogeneous graph high-order neighbor convolutional network, and the specific steps are as follows:

[0037] (1) Obtain a network graph dataset. The network graph is an undirected graph in which nodes have different attributes, which are represented by the node feature matrix X.

[0038] (2) Adaptively combine the first-order to k-order neighbor information to learn a new adjacency matrix based on the mixing function f

[0039] (3) Use the multi-layer perceptron MLP to generate the soft label matrix B of the nodes, and calculate the class similarity matrix S between the nodes based on it.

[0040] (4) Using the graph adjacency matrix of mixed multi-order neighbors The node features are iteratively updated with the node class similarity matrix S to obtain Where n is the number of nodes and c is the number of node categories.

[0041] (5) Use the softmax function to embed the generated nodes Make node category prediction.

[0042] In step (2), the new adjacency matrix is ​​obtained This is achieved through the following formula:

[0043]

[0044] in is the relearned graph adjacency matrix, represents the learnable parameter matrix, Θ i T means Θ i The transpose of A i represents the i-order neighbor information matrix, when i=0, A 0 represents the identity matrix, k represents the longest neighbor class, and the value range of k can be set between 1 and 10 according to the task requirements, depending on the task requirements and computing resources. i! refers to the factorial of i, and f represents the mixing function. Different parameters Θ are learned by minimizing the loss function i ,This node classification method dynamically adjusts the adjacency matrix according to the specific ,structure of the graph and the task requirements.

[0045] In the step (3), it is necessary to calculate the class similarity matrix between nodes, and its workflow includes:

[0046] The node feature matrix X of the graph is input into a multi-layer perceptron MLP and normalized with the softmax function to ensure that the sum of the probabilities of each node's soft label in all categories is 1 to generate the predicted category probability of the node and obtain the final soft label matrix This is achieved through the following formula:

[0047] B = softmax(σ(MLP(X)))

[0048] Among them, X represents the node feature matrix, σ represents the activation function, and B represents the soft label matrix.

[0049] Using the obtained soft label matrix B, we further calculate the class similarity matrix between nodes It is expressed by the following formula:

[0050] S=BB T

[0051] Among them, the element S of S ij Measuring the node v i and node v j The probability of belonging to the same category.

[0052] The step (4) is achieved by the following formula:

[0053]

[0054] Among them, Z (l) is the node feature representation Z at layer l (0)=X, and is the weight matrix of the l-1th layer, express The degree matrix of Z is calculated through six layers in the present invention. (6) Embed Z as the final node.

[0055] The step (5) generates node embedding in the model After that, the softmax function is used to predict the category of the node. This is achieved through the following formula:

[0056]

[0057] in, is the soft label of the prediction result,

[0058] To ensure the reliability of the node class similarity matrix S, a multi-layer perceptron (MLP) is pre-trained and the cross entropy loss function is used to optimize the parameters of the MLP. The loss function is:

[0059]

[0060] in, is node v i The predicted label, y i ∈{0,1} 1×c is the true label, is the training set, and loss(·) is the cross entropy loss function.

[0061] In addition, the pre-trained MLP layer is fine-tuned to further optimize the model performance.

[0062] The loss function of GCN is:

[0063]

[0064] in, is node v i The eigenvector of i ∈{0,1} 1×c is the true label, is the training set, and loss(·) is the cross entropy loss function.

[0065] The loss function of this node classification method is designed as:

[0066]

[0067] Among them, γ is a hyperparameter used to balance the two loss terms.

[0068] To demonstrate the effectiveness of the present invention, the mixed heterogeneous graph high-order neighbor convolutional network model of the present invention is compared with the following seven common models to analyze the performance of the classification task.

[0069] As shown in Table 1, the table shows that the node classification performance of the present invention is compared with H2GCN, GPR-GNN, CPGNN-MLP, CPGNN-Cheby, BM-GCN, and HOG-GCN on two real data sets. We select node classification accuracy for performance evaluation and record the average and standard deviation of 10 tests performed on each model on each data set as the final result. The tabular data shows the node classification performance of each model after running on the two data sets, where "-" indicates that the comparative experimental network model did not perform the node classification task on the data set.

[0070] Table 1 Comparison of Accuracy clustering indicators (mean) on different image datasets

[0071]

[0072] As can be seen from Table 1, the performance of the mixed heterogeneous graph high-order neighbor convolutional network is better than or close to other models. This shows that the mixed heterogeneous graph high-order neighbor convolutional network can effectively capture the complex relationship between nodes when processing node classification tasks, especially in heterogeneous graphs.

[0073] Based on the same inventive concept, the embodiment of the present invention discloses a citation network node classification system of a hybrid heterogeneous graph high-order neighbor convolution network, including: a data acquisition module for acquiring a citation network graph data set, the citation network graph is an undirected graph, and the nodes have different attributes, which are expressed as a node feature matrix X; a matrix update and convolution module, which uses a hybrid function to adaptively combine first-order to k-order neighbor information to learn a new adjacency matrix Use the multi-layer perceptron MLP to generate the soft label matrix of the node and calculate the class similarity matrix S between the nodes; use the graph adjacency matrix of the mixed multi-order neighbors The node features are iteratively updated with the node class similarity matrix S to obtain the node embedding Prediction module, embedding the generated nodes Make node category prediction.

[0074] Based on the same inventive concept, an embodiment of the present invention discloses a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the citation network node classification method of the hybrid heterogeneous graph high-order neighbor convolutional network.

Claims

1. A citation network node classification method based on a hybrid heterogeneous graph high-order neighbor convolutional network, characterized in that: The steps include: (1) Obtain a network graph dataset. The network graph is an undirected graph with nodes having different attributes, which is represented by a node feature matrix X. (2) Adaptively combine first-order to k-order neighbor information to learn a new adjacency matrix based on a mixing function (3) Use the multi-layer perceptron MLP to generate the node soft label matrix B, and calculate the class similarity matrix S between nodes based on it; (4) Using the graph adjacency matrix of mixed multi-order neighbors Iteratively update the node features with the node class similarity matrix S to obtain the node embedding Where n is the number of nodes, c is the number of node classifications; (5) Embed the generated nodes Make node category prediction.

2. The citation network node classification method of the mixed heterogeneous graph high-order neighbor convolutional network according to claim 1 is characterized in that: In step (2), the new adjacency matrix is ​​obtained This is achieved through the following formula: in is the relearned graph adjacency matrix, represents the learnable parameter matrix, Θ i T means Θ i The transpose of A i represents the i-order neighbor information matrix, A represents the neighbor information matrix, when i=0, A 0 represents the identity matrix, k represents the longest neighbor class, i! refers to the factorial of i, and f represents the mixing function.

3. The citation network node classification method of the mixed heterogeneous graph high-order neighbor convolutional network according to claim 1 is characterized in that: The workflow of calculating the class similarity matrix between nodes in step (3) includes: The node feature matrix X of the graph is input into a pre-trained multi-layer perceptron MLP and normalized with the softmax function to ensure that the sum of the probabilities of each node's soft label in all categories is 1 to generate the predicted category probability of the node and obtain the final soft label matrix Using the obtained soft label matrix B, we further calculate the class similarity matrix between nodes 4. The citation network node classification method of the mixed heterogeneous graph high-order neighbor convolutional network according to claim 1 is characterized in that: The node embedding Z in step (4) is implemented by the following formula: Among them, Z (l) is the node feature representation of the lth layer, where Z (0) =X, and is the weight matrix of the l-1th layer, express The degree matrix of .

5. The citation network node classification method of the mixed heterogeneous graph high-order neighbor convolutional network according to claim 1 is characterized in that: In step (5), the generated nodes are embedded The softmax function is used to predict the node category.

6. The citation network node classification method of the mixed heterogeneous graph high-order neighbor convolutional network according to claim 1 is characterized in that: By minimizing the loss function The parameters in the node classification method are obtained by training, where γ is a hyperparameter used to balance the two loss terms. is node v i The predicted label, y i ∈{0,1} 1×c is the true label, is node v i The characteristic vector of is the training set, and loss(·) is the cross entropy loss function.

7. A citation network node classification system based on a hybrid heterogeneous graph high-order neighbor convolutional network, characterized in that: include: The data collection module is used to obtain the citation network graph dataset. The citation network graph is an undirected graph with different nodes, which is represented by the node feature matrix X. Matrix update and convolution module, using a mixing function to adaptively combine first-order to k-order neighbor information to learn a new adjacency matrix Use the multi-layer perceptron MLP to generate the soft label matrix of the node and calculate the class similarity matrix S between the nodes; use the graph adjacency matrix of the mixed multi-order neighbors The node features are iteratively updated with the node class similarity matrix S to obtain the node embedding Where n is the number of nodes, c is the number of node classifications; Prediction module, embedding the generated nodes Make node category prediction.

8. The citation network node classification system of the hybrid heterogeneous graph high-order neighbor convolutional network according to claim 7, characterized in that: In the matrix update and convolution module, the new adjacency matrix This is achieved through the following formula: in is the relearned graph adjacency matrix, represents the learnable parameter matrix, Θ i T means Θ i The transpose of A i represents the i-order neighbor information matrix, when i=0, A 0 represents the identity matrix, k represents the longest neighbor class, i! refers to the factorial of i, and f represents the mixing function.

9. The citation network node classification system of the hybrid heterogeneous graph high-order neighbor convolutional network according to claim 7, characterized in that: In the matrix update and convolution module, the node embedding Z is implemented by the following formula: Among them, Z (l) is the node feature representation of the lth layer, where Z (0) =X, and is the weight matrix of the l-1th layer, express The degree matrix of .

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method for classifying citation network nodes of a mixed heterogeneous graph high-order neighbor convolutional network according to any one of claims 1 to 6 are implemented.

Images