GCN and multi-classifier fused citation network node classification method and classification system

Abstract

The invention provides a citation network node classification method and system fusing GCN and multiple classifiers, and the method comprises the following steps: enabling a whole citation network to serve as an undirected graph G comprising node attributes and a topological structure; calculating an adjacent matrix and a feature matrix of the undirected graph G; initializing an edge weight matrix of the undirected graph G, and calculating an edge weight of the undirected graph G according to the feature matrix; the calculated edge weights are converged to nodes at the two ends of the edge, and the feature matrix is updated; filtering and amplifying the features in the updated feature matrix; and respectively inputting the adjacent matrix and the updated feature matrix into n classifiers for prediction, and fusing prediction results to obtain a final classification result. According to the classification method, effective semantic information in the network is comprehensively considered, prediction is carried out through a plurality of classifiers, fusion is carried out by using fuzzy integrals, and a more accurate and stable classification effect can be obtained.

Description

Technical Field

[0001] The present invention relates to the field of graph networks and deep learning technology, and mainly to a citation network node classification method and classification system integrating GCN and multiple classifiers, aiming to improve the classification accuracy and stability of citation network nodes by integrating feature information of different levels and dimensions. Background Art

[0002] With the significant progress in graph data research, improving the accuracy of node classification has become one of the hot issues in academia and industry. Researchers have proposed a variety of innovative models and algorithms by deeply exploring the characteristics and structure of graph data, which has promoted the development of practical problem solving and artificial intelligence technology. Traditional machine learning methods, such as support vector machines, decision trees, and random forests, usually rely on numerical feature representations and corresponding labels from literature for training. These methods are insufficient when processing complex graph structure data.

[0003] In recent years, deep learning-based node classification methods have become more common due to their superiority in processing graph-structured data. Graph Convolutional Neural Networks (GCN) and its variants such as Graph Attention Networks (GAT) and GraphSAGE, etc., learn node feature representations through deep network structures, significantly improving the performance of classification tasks. These methods use the convolution operation of the graph Laplacian matrix and the aggregation of neighbor node features to update node representations, enhancing the model's ability to capture local and global information.

[0004] Although existing deep learning models have made progress in node classification, they still have limitations in capturing global information and aggregating information. Therefore, it is urgent to propose a more efficient aggregation strategy and graph convolution operation to make full use of the edge weights and attribute information between nodes to improve classification accuracy and model generalization ability. Summary of the invention

[0005] In view of this, the present invention provides a citation network node classification method and classification system integrating GCN and multiple classifiers to solve the problems existing in the prior art.

[0006] On one hand, the present invention proposes a citation network node classification method integrating GCN and multiple classifiers, comprising:

[0007] S1: Based on the knowledge citation structure of the academic field, the entire citation network is regarded as an undirected graph G including node attributes and topological structure, in which each academic paper is regarded as a node in the network, and the citation relationship between documents is regarded as the edge between nodes;

[0008] S2: Calculate the adjacency matrix and feature matrix X of the undirected graph G, where: N is the number of nodes, C is the number of features of each node, the C features represent whether the C words in the feature word dictionary appear in the citation, and the i-th row in the feature matrix X represents the feature vector of the i-th node;

[0009] S3: Initialize the edge weight matrix of the undirected graph G, and calculate the edge weight of the undirected graph G according to the feature matrix;

[0010] S4: Aggregate the edge weights calculated in S3 to the nodes at both ends of the edge, and update the feature matrix X;

[0011] S5: filtering and amplifying the features in the updated feature matrix obtained in S4;

[0012] S6: Input the adjacency matrix obtained in S2 and the feature matrix obtained in S5 into n classifiers respectively, and obtain n first classification result matrices M corresponding to the n classifiers one by one, where n is greater than 1, and the value of the i-th row and j-th column in the first classification result matrix M corresponding to each classifier represents the predicted value that the i-th node belongs to the j-th category;

[0013] The n classifiers all use a GCN model with the same structure, and each GCN model is trained independently to obtain n trained GCN models;

[0014] S7: The n first classification result matrices M obtained in S6 are fused by fuzzy integral to obtain the final classification result matrix Out, where: N is the number of nodes, Y is the number of categories, where the value of the i-th row and j-th column in the final classification result matrix Out represents the predicted value that the i-th node belongs to the j-th category;

[0015] S8: Select the category corresponding to the largest value in each row from the final classification result matrix obtained in S7 as the classification result of the corresponding node, that is, to achieve the classification of the paper.

[0016] Preferably, in S3, any two nodes d in the undirected graph G 1 and d 2 The edge weight w(d 1 ,d 2 ) is calculated as:

[0017]

[0018] Where: F 1 and F 2 They are node d 1 and d 2 The feature word set of |F 1 ∩F 2 | represents node d 1and d 2 The size of the intersection of the feature word sets, |F 1 ∪F 2 | represents node d 1 and d 2 The size of the union of the feature word sets.

[0019] Further preferably, in S4, the feature matrix X is updated using the following formula:

[0020]

[0021] In the formula, represents the i-th row of the updated feature matrix X, N(i) represents the set of adjacent nodes of the i-th node, w(i,j) represents the edge weight between the i-th node and the j-th node, X j Represents the j-th row of the feature matrix X before update.

[0022] Further preferably, in S5, the steps of filtering and amplifying are as follows:

[0023] First, calculate the variance of each feature;

[0024] Then, define a threshold θ var , retain features whose variance is greater than the threshold;

[0025] Finally, it will be greater than the preset threshold θ f The eigenvalues ​​are increased by the corresponding multiples according to the importance index to amplify the obvious features.

[0026] Further preferably, S5 further includes a step of standardizing the amplified features so that they have a distribution with a mean of 0 and a standard deviation of 1.

[0027] Further preferably, in S6, n trained GCN models are obtained through different random initializations and randomness of input data order.

[0028] More preferably, in S7, Sugeno fuzzy integral is used for fusion.

[0029] Further preferably, in S7, fusing the n first classification result matrices M by fuzzy integral comprises the following steps:

[0030] S71: For each category, extract corresponding columns from the n first classification result matrices M corresponding to the n classifiers to form Y second classification result matrices Z corresponding to each category, where: N is the number of nodes, n represents the number of classifiers, wherein the second classification result matrix Z(k) corresponding to the k-th category is obtained as follows: extract the k-th column in the first classification result matrix M corresponding to each classifier respectively to form the second classification result matrix Z(k) corresponding to the k-th category;

[0031] S72: The second classification result matrix Z corresponding to each category is taken as input in turn, and is calculated by Sugeno fuzzy integral function to obtain the fusion result of each category, wherein the calculation formula of the fusion result of the kth category is as follows:

[0032]

[0033] Among them, Fused(k) represents the fusion result of the k-th category, μ(i) is the fuzzy measure, which represents the weight or credibility corresponding to the i-th classifier, and Z(k) i represents the i-th column in the second classification result matrix Z(k) corresponding to the k-th category;

[0034] S73: Combine the fusion results of each category to obtain the final classification result matrix

[0035] The present invention also proposes a citation network node classification system integrating GCN and multiple classifiers, which is used to execute the above-mentioned citation network node classification method integrating GCN and multiple classifiers.

[0036] The citation network node classification method and classification system provided by the present invention, which integrates GCN and multiple classifiers, re-integrates feature information of different levels and dimensions, integrates, filters and screens low-dimensional edge weights with high-dimensional node features, and uses multiple classifiers to obtain multiple prediction results and fuses them through fuzzy integrals, which can improve the overall classification performance and reduce the error caused by a single classifier to a certain extent. Compared with the mainstream classification methods in the current graph neural network research, the classification results of the present invention are improved in terms of accuracy, recall rate and other indicators, significantly improving the accuracy and stability of node classification, and providing an innovative and efficient solution for node classification tasks in graph networks. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further explained below in conjunction with specific embodiments, but the present invention is not limited thereto.

[0038] In order to improve the comprehensive effect of citation node classification, this implementation plan starts from the fusion of features at different levels and dimensions, considers the various effective semantic information in the network, and proposes a citation network node classification method that integrates GCN and multiple classifiers. The specific steps are as follows:

[0039] S1: Based on the knowledge citation structure of the academic field, the entire citation network is regarded as an undirected graph G including node attributes and topological structure, in which each academic paper is regarded as a node in the network, and the citation relationship between documents is regarded as the edge between nodes. When one paper cites another paper, it means that an edge is established between the two papers.

[0040] By converting the knowledge citation structure of an academic field into a network graph, the nodes in the network graph can be studied using graph theory and network analysis methods;

[0041] S2: Calculate the adjacency matrix and feature matrix X of the undirected graph G, where: N is the number of nodes, C is the number of features of each node, the C features represent whether the C words in the feature word dictionary appear in the citation, and the i-th row in the feature matrix X represents the feature vector of the i-th node;

[0042] S3: Initialize the edge weight matrix of the undirected graph G, and calculate the edge weight of the undirected graph G according to the feature matrix;

[0043] Among them, any two nodes d in the undirected graph G 1 and d 2 The edge weight w(d 1 ,d 2 ) is calculated as:

[0044]

[0045] Where: F 1 and F 2 They are node d 1 and d 2 The feature word set of |F 1 ∩F 2 | represents node d 1 and d 2 The size of the intersection of the feature word sets, |F 1 ∪F 2 | represents node d 1 and d 2 The size of the union of the feature word sets.

[0046] The edge weight between two documents depends on the ratio of the number of common feature words they contain to the total number of feature words they contain C. If the two documents have more common feature words, the edge weight between them will be higher, that is, the two nodes have greater similarity, and vice versa.

[0047] S4: Aggregate the edge weights calculated in S3 to the nodes at both ends of the edge, update the feature matrix X, and obtain an updated feature matrix;

[0048] Due to the characteristics of the citation network, when a feature in a node is zero, it means that the word does not appear in the paper node. Therefore, when updating the node feature according to the edge weight, the feature should be skipped. The i-th row of the updated feature matrix X (i.e., the updated feature vector of the i-th node) ) is expressed as:

[0049]

[0050] In the formula, N(i) represents the set of adjacent nodes of the i-th node, w(i,j) represents the edge weight between the i-th node and the j-th node, and X j Represents the j-th row of the feature matrix X before update (i.e., the feature vector of the j-th node);

[0051] S5: To further improve the model performance, the features in the updated feature matrix obtained in S4 are filtered and amplified;

[0052] Through filtering and amplification, it is possible to filter out unclear features and amplify obvious features, thereby achieving the purpose of removing noise, highlighting important features, and cleaning low-correlation features.

[0053] The steps of filtering and amplification are as follows:

[0054] First, calculate the variance of each feature;

[0055] Then, define a threshold θ var , retain features whose variance is greater than the threshold;

[0056] Finally, it will be greater than the preset threshold θ f The eigenvalues ​​are increased by the corresponding multiples according to the importance index to amplify the obvious features.

[0057] As an improvement of the technical solution, S5 also includes the following step: standardizing the amplified features so that they have a distribution with a mean of 0 and a standard deviation of 1.

[0058] After the above aggregation operation, the feature matrix X corresponding to the citation network undirected graph G can simultaneously contain the semantic information of the edges and neighbor nodes in the graph, and filter and amplify the attribute values ​​in the feature matrix, thereby preliminarily improving the performance of the model in the preprocessing stage;

[0059] S6: Input the adjacency matrix obtained in S2 and the feature matrix obtained in S5 into n classifiers respectively, and obtain n first classification result matrices M corresponding to the n classifiers one by one, where n is greater than 1, and the value of the i-th row and j-th column in the first classification result matrix M corresponding to each classifier represents the predicted value that the i-th node belongs to the j-th category;

[0060] The n classifiers use a GCN model with the same structure, and each GCN model is trained independently to obtain n trained GCN models, wherein the n trained GCN models are obtained through different random initializations and randomness of the order of input data;

[0061] S7: The n first classification result matrices M obtained in S6 are fused by fuzzy integral to obtain the final classification result matrix Out, where: N is the number of nodes, Y is the number of categories, where the value of the i-th row and j-th column in the final classification result matrix Out represents the predicted value that the i-th node belongs to the j-th category;

[0062] Among them, fuzzy integral is a method of integrating input function and fuzzy measure, which can handle uncertainty in fuzzy environment. In the present invention, the mathematical basis of fusion is Sugeno fuzzy integral. Sugeno fuzzy integral is a commonly used fuzzy integral method, which calculates the output value based on the value of fuzzy measure and input function, can effectively fuse multiple information sources, handle uncertainty and fuzziness, and is widely used in multi-classifier fusion and decision-making process.

[0063] Preferably, fusing n first classification result matrices M by fuzzy integral comprises the following steps:

[0064] S71: For each category, extract corresponding columns from the n first classification result matrices M corresponding to the n classifiers to form Y second classification result matrices Z corresponding to each category, where: N is the number of nodes, n represents the number of classifiers, wherein the second classification result matrix Z(k) corresponding to the k-th category is obtained as follows: extract the k-th column in the first classification result matrix M corresponding to each classifier respectively to form the second classification result matrix Z(k) corresponding to the k-th category;

[0065] S72: The second classification result matrix Z corresponding to each category is taken as input in turn, and is calculated by Sugeno fuzzy integral function to obtain the fusion result of each category, wherein the calculation formula of the fusion result of the kth category is as follows:

[0066]

[0067] Among them, Fused(k) represents the fusion result of the k-th category, which is a vector, μ(i) is the fuzzy measure, which represents the weight or credibility corresponding to the i-th classifier, and Z(k) i represents the i-th column in the second classification result matrix Z(k) corresponding to the k-th category;

[0068] S73: Combine the fusion results of each category to obtain the final classification result matrix

[0069] S8: Select the category corresponding to the largest value in each row from the final classification result matrix obtained in S7 as the classification result of the corresponding node, that is, to achieve the classification of the paper.

[0070] The present invention also proposes a citation network node classification system integrating GCN and multiple classifiers, which is used to execute the above-mentioned citation network node classification method integrating GCN and multiple classifiers.

[0071] The present invention integrates the citation network node classification method of GCN and multiple classifiers, that is, the classification system, which re-integrates feature information of different levels and dimensions, fuses, filters and amplifies low-dimensional edge weights with high-dimensional node features, and uses multiple classifiers to obtain multiple prediction results and fuses them through fuzzy integrals, which can improve the overall classification performance and reduce the error caused by a single classifier to a certain extent. Compared with the mainstream classification methods in the current graph neural network research, the classification results of the present invention are improved in terms of accuracy, recall rate and other indicators, significantly improving the accuracy and stability of node classification, and providing an innovative and efficient solution for node classification tasks in graph networks.

[0072] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Python and literal scripting language JavaScript, etc.

[0073] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not disclosed by the present invention. The specification and examples are to be considered exemplary only, and the true scope and spirit of the present invention are indicated by the following claims.

[0074] It should be understood that the present invention is not limited to what has been described above and that various modifications and changes may be made without departing from its scope. The scope of the present invention is limited only by the appended claims.

Claims

1. A citation network node classification method integrating GCN and multiple classifiers, characterized in that: include: S1: Based on the knowledge citation structure of the academic field, the entire citation network is regarded as an undirected graph G including node attributes and topological structure, in which each academic paper is regarded as a node in the network, and the citation relationship between documents is regarded as the edge between nodes; S2: Calculate the adjacency matrix and feature matrix X of the undirected graph G, where: N is the number of nodes, C is the number of features of each node, the C features represent whether the C words in the feature word dictionary appear in the citation, and the i-th row in the feature matrix X represents the feature vector of the i-th node; S3: Initialize the edge weight matrix of the undirected graph G, and calculate the edge weight of the undirected graph G according to the feature matrix; S4: Aggregate the edge weights calculated in S3 to the nodes at both ends of the edge, and update the feature matrix X; S5: filtering and amplifying the features in the updated feature matrix obtained in S4; S6: Input the adjacency matrix obtained in S2 and the feature matrix obtained in S5 into n classifiers respectively, and obtain n first classification result matrices M corresponding to the n classifiers one by one, where n is greater than 1, and the value of the i-th row and j-th column in the first classification result matrix M corresponding to each classifier represents the predicted value that the i-th node belongs to the j-th category; The n classifiers all use a GCN model with the same structure, and each GCN model is trained independently to obtain n trained GCN models; S7: The n first classification result matrices M obtained in S6 are fused by fuzzy integral to obtain the final classification result matrix Out, where: N is the number of nodes, Y is the number of categories, and the value of the i-th row and j-th column in the final classification result matrix Out represents the predicted value that the i-th node belongs to the j-th category; S8: Select the category corresponding to the largest value in each row from the final classification result matrix obtained in S7 as the classification result of the corresponding node, that is, to achieve the classification of the paper.

2. The citation network node classification method integrating GCN and multiple classifiers according to claim 1 is characterized by: In S3, the edge weight w(d1, d2) between any two nodes d1 and d2 in the undirected graph G is calculated as: Where: F1 and F2 are the feature word sets of nodes d1 and d2 respectively, |F1∩F2| represents the size of the intersection of the feature word sets of nodes d1 and d2, and |F1∪F2| represents the size of the union of the feature word sets of nodes d1 and d2.

3. The citation network node classification method integrating GCN and multiple classifiers according to claim 1 is characterized by: In S4, the feature matrix X is updated using the following formula: In the formula, represents the i-th row of the updated feature matrix X, N(i) represents the set of adjacent nodes of the i-th node, w(i,j) represents the edge weight between the i-th node and the j-th node, X j Represents the j-th row of the feature matrix X before update.

4. The citation network node classification method integrating GCN and multiple classifiers according to claim 1 is characterized by: In S5, the filtering and amplification steps are as follows: First, calculate the variance of each feature; Then, define a threshold θ var , retain features whose variance is greater than the threshold; Finally, it will be greater than the preset threshold θ f The eigenvalues ​​are increased by the corresponding multiples according to the importance index to amplify the obvious features.

5. The citation network node classification method integrating GCN and multiple classifiers according to claim 1, characterized in that: S5 also includes a step of standardizing the amplified features so that the features have a distribution with a mean of 0 and a standard deviation of 1.

6. The citation network node classification method integrating GCN and multiple classifiers according to claim 1, characterized in that: In S6, n trained GCN models are obtained with different random initializations and randomness of the order of input data.

7. The citation network node classification method integrating GCN and multiple classifiers according to claim 1 is characterized by: In S7, Sugeno fuzzy integral is used for fusion.

8. The citation network node classification method integrating GCN and multiple classifiers according to claim 1 is characterized by: In S7, fusing n first classification result matrices M by fuzzy integral includes the following steps: S71: For each category, extract corresponding columns from the n first classification result matrices M corresponding to the n classifiers to form Y second classification result matrices Z corresponding to each category, where: N is the number of nodes, n represents the number of classifiers, wherein the second classification result matrix Z(k) corresponding to the k-th category is obtained as follows: extract the k-th column in the first classification result matrix M corresponding to each classifier respectively to form the second classification result matrix Z(k) corresponding to the k-th category; S72: The second classification result matrix Z corresponding to each category is taken as input in turn, and is calculated by Sugeno fuzzy integral function to obtain the fusion result of each category, wherein the calculation formula of the fusion result of the kth category is as follows: Among them, Fused(k) represents the fusion result of the k-th category, μ(i) is the fuzzy measure, which represents the weight or credibility corresponding to the i-th classifier, and Z(k) i represents the i-th column in the second classification result matrix Z(k) corresponding to the k-th category; S73: Combine the fusion results of each category to obtain the final classification result matrix 9. A citation network node classification system integrating GCN and multiple classifiers, characterized by: A citation network node classification method for implementing the fusion GCN and multiple classifiers described in any one of claims 1-8.

Images