A method for academic network paper classification based on graph contrastive learning

Abstract

The present invention provides an academic network paper classification method based on graph contrast learning. The method uses meta-path to model the academic network as paper sub-graphs with different semantics, and comprehensively considers different types of node information in the meta-path, reconstructs paper node features to construct a meta-path view of the paper sub-graph, can learn more comprehensive paper node features, improve classification accuracy, set different data enhancement strategies for large-degree paper nodes and small-degree paper nodes respectively to eliminate the long-tail distribution problem of paper interaction data, thereby constructing an enhanced view, which is suitable for various long-tail distribution scenarios, and uses a graph contrast learning method to perform contrast learning on the embedding vectors of the meta-path view and the enhanced view, thereby obtaining a more comprehensive and effective paper embedding vector, and classifies the paper nodes through a multi-layer perceptron, thereby finally improving the accuracy of academic network paper classification.

Description

Technical Field

[0001] The present invention relates to the technical field of academic network data processing, and in particular to an academic network paper classification method based on graph contrast learning. Background Art

[0002] The academic network not only contains a wealth of information on papers, conferences, and scholars, but also records the citation relationships between academic papers and the relationships between scholars writing papers. Studying the classification of papers in the academic network helps to organize a large number of academic papers according to themes and fields, so that researchers can more effectively retrieve papers on specific themes or fields, thereby improving research efficiency.

[0003] At present, for academic networks with multiple node types and multiple relationship types, most paper classification methods model academic networks according to the set meta-path, but such methods ignore the information of intermediate nodes in the meta-path, such as the information of authors and conference nodes, which will cause the loss of information represented by paper nodes to a certain extent. The paper classification method based on graph contrastive learning does not rely on the label information of the paper and is a feasible approach. Many related studies have achieved good results. However, most existing studies have ignored the long-tail distribution problem of the interaction data between papers in academic networks, such as a few papers have a high number of citations, while most papers have a relatively low number of citations. This long-tail distribution problem of paper interaction data will lead to an imbalance in the information learned by paper nodes in the model, resulting in inaccurate paper classification results. Summary of the invention

[0004] To this end, the purpose of the present invention is to provide an academic network paper classification method based on graph contrast learning to improve the accuracy of paper classification results.

[0005] A method for classifying academic network papers based on graph contrastive learning, including:

[0006] Step S1, obtaining an academic network data set consisting of papers to be classified on an academic resource platform, and constructing an initial feature vector of a node, wherein the initial feature vector includes: an initial feature vector of a paper node constructed by counting keywords of the paper, and an initial feature vector of an author and conference node obtained by random initialization;

[0007] Step S2, constructing paper subgraphs with different semantics in the academic network based on meta-paths;

[0008] Step S3, constructing a meta-path view based on the paper subgraph and the initial feature vector, learning the embedding vector of the paper node in the meta-path view, fusing the embedding vectors of the meta-path view under different semantics, and projecting through a multi-layer perceptron to obtain the final embedding vector under the meta-path view;

[0009] Step S4, constructing an enhanced view based on the paper subgraph and the initial feature vector, learning the embedding vector of the paper node in the enhanced view, fusing the enhanced view embedding vectors under different semantics, and projecting through a multi-layer perceptron to obtain the final embedding vector under the enhanced view;

[0010] Step S5, constructing a loss function of the meta-path view and a loss function of the enhanced view respectively, and performing weighted fusion on the loss function of the meta-path view and the loss function of the enhanced view to obtain a total contrast loss function, optimizing the final embedding vector under the meta-path view and the final embedding vector under the enhanced view by the total contrast loss function, and obtaining an optimized final embedding vector under the meta-path view and an optimized final embedding vector under the enhanced view;

[0011] Step S6, concatenate the final embedding vector under the optimized meta-path view and the final embedding vector under the optimized enhancement view to obtain a concatenated embedding representation, and input the concatenated embedding representation into the paper classification model to obtain a classification result.

[0012] Beneficial effects of the present invention:

[0013] This method uses meta-paths to model academic networks as paper subgraphs with different semantics. At the same time, it comprehensively considers the information of different types of nodes in the meta-path, reconstructs the features of paper nodes, and constructs a meta-path view of the paper subgraph. It can learn more comprehensive paper node features and improve the classification accuracy. Different data enhancement strategies are set for large-degree paper nodes and small-degree paper nodes to eliminate the long-tail distribution problem of paper interaction data, thereby constructing an enhanced view that is suitable for various long-tail distribution scenarios. The graph contrast learning method is used to compare the embedding vectors of the meta-path view and the enhanced view, and a more comprehensive and effective paper embedding vector is obtained. The paper nodes are classified through a multi-layer perceptron, which ultimately improves the accuracy of academic network paper classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Picture 1 It is a flow chart of the academic network paper classification method based on graph contrast learning provided by the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] See also Picture 1 The present invention provides an academic network paper classification method based on graph contrast learning, comprising steps S1 to S6:

[0017] Step S1, obtain an academic network data set consisting of papers to be classified on an academic resource platform, and construct an initial feature vector of a node. The initial feature vector includes: an initial feature vector of a paper node constructed by counting the keywords of the paper, and an initial feature vector of an author and conference node obtained by random initialization.

[0018] Specifically, the academic network dataset includes the keywords of papers, the association information between authors and papers, papers and papers, and papers and conferences. The initial feature vector of paper nodes is constructed by counting the keywords of papers, and the initial feature vectors of authors and conference nodes are obtained by random initialization. H.G. for H.G. =( V , E , T , R ),in, V is a node set, E is a set of relations, T is a node type collection, R is a set of relation types that satisfies , that is, the sum of the number of node types and the number of relationship types exceeds 2. The degree of the paper nodes in the academic network is counted and sorted from large to small. The top 10% of the degree ranking is divided into high-degree paper nodes, and the bottom 40% of the degree ranking is divided into low-degree paper nodes. The low-degree paper node is defined as { node 1,…, node s},in node 1 is the first small degree paper node, node s is the sth small degree paper node, and the large degree paper node is defined as { Node 1,…, Node m},in Node 1 is the first large paper node, Node m is the mth large degree paper node.

[0019] Step S2: construct paper subgraphs with different semantics in the academic network based on meta-paths.

[0020] Among them, in step S2: determine the type of paper node to be classified, design a set of symmetric meta-paths based on the association information between authors and papers, papers and papers, and papers and conferences. Different meta-paths represent different semantics. The kth symmetric meta-path for: , where the source node type and the terminal node type The same and all are paper nodes to be classified, For the second node type, , and They are the first edge type, the second edge type and the Type of edge connection.

[0021] Step S2 specifically includes: starting from the kth element path The source node of the academic network starts to traverse, and the nodes of the source node type that conform to the meta-path pattern are retained until the traversal ends, forming the kth paper subgraph For example, for a meta-path that satisfies the “paper-author-paper” pattern, the final constructed paper subgraph only contains paper node types, and the direct connections between papers are established through authors. H.G. Divided into n semantic paper subgraphs { G 1,…, G n},in G 1 is the first paper subgraph, G n is the nth paper subgraph.

[0022] Step S3, constructing a meta-path view based on the paper subgraph and the initial feature vector, learning the embedding vector of the paper node in the meta-path view, fusing the meta-path view embedding vectors under different semantics, and projecting through a multi-layer perceptron to obtain the final embedding vector under the meta-path view.

[0023] Wherein, step S3 specifically includes:

[0024] Step S31, mapping the initial feature vector of the paper node and the initial feature vectors of the author and conference nodes to the same vector space through a multi-layer perceptron, reconstructing the node features in the paper subgraph, and thus constructing a meta-path view;

[0025] In step S31, for the kth element path , add the intermediate node of the meta-path as a virtual node to the kth paper subgraph G k For example, the author node or conference node, and according to the connection mode of the meta-path, the corresponding paper node is connected to obtain the k-th reconstructed subgraph , using graph convolutional encoder to learn paper nodes in The reconstructed feature vector is used as the feature vector of the meta-path view.

[0026] The construction process of the feature vector in the meta-path view is as follows:

[0027] ;

[0028] in, To reconstruct the subgraph after adding virtual nodes, for The adjacency matrix of the node is determined by the connection between the nodes. If there is an edge, it is 1, otherwise it is 0. for The feature vector in . Meta path view In this process, the virtual node is only used to reconstruct the feature information of the target node and will not be used again in the future. It is a graph convolution encoder. The specific process is as follows:

[0029] ;

[0030] in, and They are the first activation function and the second activation function in the graph convolutional network, W 1 and W 2 are the first weight matrix and the second weight matrix in the graph convolutional network, b 1 and b 2 are the first bias vector and the second bias vector in the graph convolutional network respectively.

[0031] Step S32, the adjacency matrix and feature vector of the meta-path view are input into the graph attention encoder to learn the embedding vector of the paper node, and the expression is as follows:

[0032] ;

[0033] in, For the meta path view The embedding vector in , It is a graph attention encoder. In the specific implementation, for the meta-path view The learning process of node i is as follows:

[0034] ;

[0035] in, For the meta path view The embedding vector of node i in the meta-path view, is the third activation function, Meta path view In the paper, the attention weight of node i about node j is: is the feature vector of paper node j, For the meta path view The set of neighbor nodes of node i.

[0036] Step S33, input the learned embedding vectors of different semantic meta-path views into the semantic level attention mechanism for semantic feature fusion, which is expressed as follows:

[0037] ;

[0038] in, is the fused paper node embedding vector under the meta-path view, is the kth meta-path view The embedding vector of for The attention weight, All nodes in share the weight, and n represents the total number of meta-path views.

[0039] Step S34: Input the fused embedding vector in step S33 into the multi-layer perceptron, project the feature into another vector space, and obtain the final embedding vector under the meta-path view. , the expression is as follows:

[0040] ;

[0041] in, f 1 is the first multi-layer perceptron, which consists of two fully connected layers and an activation function.

[0042] Step S4, constructing an enhanced view based on the paper subgraph and the initial feature vector, learning the embedding vector of the paper node in the enhanced view, fusing the enhanced view embedding vectors under different semantics, and projecting through a multi-layer perceptron to obtain the final embedding vector under the enhanced view.

[0043] Wherein, step S4 specifically includes:

[0044] Step S41, constructing an enhanced view by setting different enhancement strategies for small-degree paper nodes and large-degree paper nodes in the paper subgraph;

[0045] Wherein, step S41 specifically includes:

[0046] Calculate the cosine similarity between the small-degree paper node and its second-order neighbors, and select the second-order neighbors with the top 10% of cosine similarity as the new first-order neighbors of the small-degree paper node; select the initial feature vector of the large-degree paper node, and use the masking mechanism to set some dimensional information of the feature vector of the large-degree paper node to 0, masking some feature information, and construct an enhanced view through the above enhancement strategy;

[0047] Among them, the cosine similarity of paper node i to paper node j is The calculation formula is:

[0048] ;

[0049] in, Subgraph for the paper G The initial eigenvector of the i-th small-degree paper node in, Subgraph for the paper G The initial eigenvector of node j in the paper, is the dot product, Indicates the length of the calculation vector;

[0050] For the large-scale paper nodes, a mask mechanism is used to cover up some feature information. The calculation formula is:

[0051] ;

[0052] in, To enhance the view, for The feature vector of the large-scale paper node i, for G The initial eigenvector of node i in the large-scale paper, M is the mask matrix.

[0053] Step S42: enhance the adjacency matrix of the view and the eigenvector Embedding vectors of different semantically enhanced views input to the graph convolutional encoder , the expression is as follows:

[0054] .

[0055] Step S43: embedding vectors of enhanced views under different semantics learned Input into the semantic level attention for semantic feature fusion to obtain the fused embedding vector , the expression is as follows:

[0056] ;

[0057] in, is the kth enhanced view The embedding vector of for The attention weight, All nodes share .

[0058] Step S44: embedding the fused vector in step S43 Input into the multi-layer perceptron, project the features into another vector space, and obtain the final embedding vector under the enhanced view , the expression is as follows:

[0059] .

[0060] Step S5, construct the loss function of the meta-path view and the loss function of the enhanced view respectively, and perform weighted fusion on the loss function of the meta-path view and the loss function of the enhanced view to obtain the total contrast loss function, and optimize the final embedding vector under the meta-path view and the final embedding vector under the enhanced view through the total contrast loss function to obtain the optimized final embedding vector under the meta-path view and the optimized final embedding vector under the enhanced view.

[0061] Wherein, step S5 specifically includes:

[0062] The nodes in the meta-path view or enhanced view are used as anchor nodes, the nodes corresponding to the anchor nodes in different views are used as positive samples, and the other nodes in different views are used as negative samples. The loss function of the meta-path view and the loss function of the enhanced view are constructed respectively. The expressions are as follows:

[0063] ;

[0064] ;

[0065] in, is the loss of paper node i under the meta-path view, is the loss of node i in the enhanced view, sim is the similarity function, is the embedding vector of paper node i under the meta-path view, is the embedding vector of paper node i under the enhanced view, is the embedding vector of paper node t under enhanced view, is the embedding vector of paper node t under the meta-path view, is the temperature coefficient;

[0066] The expression of the total contrast loss function is:

[0067] ;

[0068] in, is the total contrast loss function, represents all training samples, is the total number of all training samples, is the weight coefficient.

[0069] Step S6, concatenate the final embedding vector under the optimized meta-path view and the final embedding vector under the optimized enhancement view to obtain a concatenated embedding representation, and input the concatenated embedding representation into the paper classification model to obtain a classification result.

[0070] In step S6, the paper classification model is a multi-layer perceptron, and the concatenated embedding representation is input into the multi-layer perceptron to calculate the classification result. , is a probability distribution matrix, each row corresponds to a sample, each column corresponds to the probability of a category, and the category with the highest probability is selected as the sample prediction result. The calculation process is:

[0071] ;

[0072] in, For splicing operation, and are the final embedding vector under the optimized meta-path view and the final embedding vector under the optimized enhanced view, respectively. f 2 is the second multi-layer perceptron, which consists of two activation functions and two fully connected layers.

[0073] In summary, according to the academic network paper classification method based on graph contrast learning provided by the present invention, the meta-path is used to model the academic network as a paper subgraph with different semantics. At the same time, the different types of node information in the meta-path are comprehensively considered, and the paper node features are reconstructed to construct a meta-path view of the paper subgraph. More comprehensive paper node features can be learned to improve the classification accuracy. Different data enhancement strategies are set for large-degree paper nodes and small-degree paper nodes to eliminate the long-tail distribution problem of paper interaction data, thereby constructing an enhanced view suitable for various long-tail distribution scenarios. The graph contrast learning method is used to compare the embedding vectors of the meta-path view and the enhanced view to obtain a more comprehensive and effective paper embedding vector. The paper nodes are classified by a multi-layer perceptron, which ultimately improves the accuracy of academic network paper classification.

[0074] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0075] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for classifying academic network papers based on graph contrastive learning, characterized in that: include: Step S1, obtaining an academic network data set consisting of papers to be classified on an academic resource platform, and constructing an initial feature vector of a node, wherein the initial feature vector includes: an initial feature vector of a paper node constructed by counting keywords of the paper, and an initial feature vector of an author and conference node obtained by random initialization; Step S2, constructing paper subgraphs with different semantics in the academic network based on meta-paths; Step S3, constructing a meta-path view based on the paper subgraph and the initial feature vector, learning the embedding vector of the paper node in the meta-path view, fusing the embedding vectors of the meta-path view under different semantics, and projecting through a multi-layer perceptron to obtain the final embedding vector under the meta-path view; Step S4, constructing an enhanced view based on the paper subgraph and the initial feature vector, learning the embedding vector of the paper node in the enhanced view, fusing the enhanced view embedding vectors under different semantics, and projecting through a multi-layer perceptron to obtain the final embedding vector under the enhanced view; Step S5, constructing a loss function of the meta-path view and a loss function of the enhanced view respectively, and performing weighted fusion on the loss function of the meta-path view and the loss function of the enhanced view to obtain a total contrast loss function, optimizing the final embedding vector under the meta-path view and the final embedding vector under the enhanced view by the total contrast loss function, and obtaining an optimized final embedding vector under the meta-path view and an optimized final embedding vector under the enhanced view; Step S6, concatenating the final embedding vector under the optimized meta-path view and the final embedding vector under the optimized enhancement view to obtain a concatenated embedding representation, and inputting the concatenated embedding representation into the paper classification model to obtain a classification result; Step S3 specifically includes: Step S31, mapping the initial feature vector of the paper node and the initial feature vectors of the author and conference nodes to the same vector space through a multi-layer perceptron, reconstructing the node features in the paper subgraph, and thus constructing a meta-path view; Step S32, input the adjacency matrix and feature vector of the meta-path view into the graph attention encoder to learn the embedding vector of the paper node; Step S33, inputting the learned embedding vectors of the meta-path views under different semantics into the semantic level attention mechanism for semantic feature fusion; Step S34, inputting the fused embedding vector in step S33 into a multi-layer perceptron, projecting the features into another vector space, and obtaining the final embedding vector under the meta-path view; Step S4 specifically includes: Step S41, constructing an enhanced view by setting different enhancement strategies for small-degree paper nodes and large-degree paper nodes in the paper subgraph, wherein the degrees of paper nodes in the academic network are counted and sorted in order from large to small, and the paper nodes with the top 10% of degree ranking are divided into large-degree paper nodes, and the paper nodes with the bottom 40% of degree ranking are divided into small-degree paper nodes. The small-degree paper nodes are defined as { node 1,…, node s },in node 1 is the first small degree paper node, node s is the sth small degree paper node, and the large degree paper node is defined as { Node 1,…, Node m },in Node 1 is the first large paper node, Node m Take the mth large-degree paper node; calculate the cosine similarity between the small-degree paper node and its second-order neighbors, and select the second-order neighbors with the top 10% of cosine similarity as the new first-order neighbors of the small-degree paper node; select the initial feature vector of the large-degree paper node, and use the masking mechanism to set some dimensional information of the feature vector of the large-degree paper node to 0, cover up some feature information, and construct an enhanced view through the above enhancement strategy; Step S42, inputting the adjacency matrix and feature vector of the enhanced view into the graph convolution encoder to learn the embedding vector of the enhanced view under different semantics; Step S43, inputting the learned embedding vectors of enhanced views under different semantics into the semantic level attention to perform semantic feature fusion; Step S44, input the fused embedding vector in step S43 into a multi-layer perceptron, project the features into another vector space, and obtain the final embedding vector under the enhanced view.

2. The method for classifying academic network papers based on graph contrastive learning according to claim 1 is characterized in that: In step S1, define the academic network dataset HG for HG =( V , E , T , R ),in, V is a node set, E is a set of relations, T is a node type collection, R is a set of relation types that satisfies , that is, the sum of the number of node types and the number of relationship types exceeds 2.

3. The method for classifying academic network papers based on graph contrast learning according to claim 2 is characterized in that: In step S2: determine the type of paper node to be classified, design a set of symmetric meta-paths, different meta-paths represent different semantics, and the kth symmetric meta-path for: , where the source node type and the terminal node type The same and all are paper nodes to be classified, For the second node type, , and They are the first edge type, the second edge type and the Types of edge connections; Step S2 specifically includes: starting from the kth element path The source node of the academic network starts to traverse, and the nodes of the source node type that conform to the meta-path pattern are retained until the traversal ends, forming the kth paper subgraph , n symmetric element paths will HG Divided into n semantic paper subgraphs { G 1,…, G n },in G 1 is the first paper subgraph, G n is the nth paper subgraph.

4. The method for classifying academic network papers based on graph contrast learning according to claim 3 is characterized in that: In step S31, for the kth element path , add the intermediate node of the meta-path as a virtual node to the kth paper subgraph G k And according to the connection mode of the meta-path, the corresponding paper node is connected to obtain the kth reconstructed subgraph , using graph convolutional encoder to learn paper nodes in The reconstructed feature vector is used as the feature vector of the meta-path view. The construction process of the feature vector in the meta-path view is as follows: ； in, To reconstruct the subgraph after adding virtual nodes, for The adjacency matrix of for The eigenvectors in Meta path view The characteristic vector of and They are the first activation function and the second activation function in the graph convolutional network, W 1 and W 2 are the first weight matrix and the second weight matrix in the graph convolutional network, b 1 and b 2 are the first bias vector and the second bias vector in the graph convolutional network respectively.

5. The method for classifying academic network papers based on graph contrast learning according to claim 4 is characterized in that: In step S41: Cosine similarity of paper node i with respect to paper node j The calculation formula is: ； in, Subgraph for the paper G The initial eigenvector of the i-th small-degree paper node in, Subgraph for the paper G The initial eigenvector of node j in the paper, is the dot product, Indicates the length of the calculation vector; For the large-scale paper nodes, a mask mechanism is used to cover up some feature information. The calculation formula is: ； in, To enhance the view, for The feature vector of node i in the large-scale paper, for G The initial eigenvector of node i in the large-scale paper, M is the mask matrix.

6. The method for classifying academic network papers based on graph contrastive learning according to claim 5 is characterized in that: Step S5 specifically includes: The nodes in the meta-path view or enhanced view are used as anchor nodes, the nodes corresponding to the anchor nodes in different views are used as positive samples, and the other nodes in different views are used as negative samples. The loss function of the meta-path view and the loss function of the enhanced view are constructed respectively. The expressions are as follows: ； ； in, is the loss of paper node i under the meta-path view, is the loss of node i in the enhanced view, sim is the similarity function, is the embedding vector of paper node i under the meta-path view, is the embedding vector of paper node i under the enhanced view, is the embedding vector of paper node t under enhanced view, is the embedding vector of paper node t under the meta-path view, is the temperature coefficient; The expression of the total contrast loss function is: ； in, is the total contrast loss function, represents all training samples, is the total number of all training samples, is the weight coefficient.

7. The method for classifying academic network papers based on graph contrastive learning according to claim 6 is characterized in that: In step S6, the paper classification model is a multi-layer perceptron, and the concatenated embedding representation is input into the multi-layer perceptron to calculate the classification result. , is a probability distribution matrix, each row corresponds to a sample, each column corresponds to the probability of a category, and the category with the highest probability is selected as the sample prediction result. The calculation process is: ； in, For splicing operation, and are the final embedding vector under the optimized meta-path view and the final embedding vector under the optimized enhanced view, respectively. f 2 is the second multi-layer perceptron, which consists of two activation functions and two fully connected layers.

Images