Paper classification method based on graph matching and self-supervised graph learning

Abstract

The invention discloses a paper classification method based on graph matching and self-supervised graph learning, and relates to the technical field of document classification based on deep learning. According to the method, literature data is represented by adopting a literature relation graph, a graph learning model ConGM based on the literature relation graph is constructed, and reference and theme association between literatures are mined through sub-graph sampling and data enhancement, linear node matching, secondary edge alignment and double-layer negative sample selection, so that precise classification of fields to which papers belong is realized.

Description

Technical Field

[0001] This invention relates to the field of document classification technology based on deep learning, and specifically to a paper classification method based on graph matching and self-supervised graph learning. Background Technology

[0002] With the ever-increasing volume of scientific research literature, effectively classifying and retrieving papers within their respective fields has become a significant challenge. Traditional paper classification methods generally employ text-based techniques, classifying papers based on keywords and textual features. However, this approach is ineffective when dealing with literature data containing complex relationships and structural information. Particularly in academic literature, complex citation relationships and thematic connections exist between different papers; relying solely on textual content for classification cannot fully utilize this structured information, thus hindering classification accuracy.

[0003] With the rise of Graph Neural Networks (GNNs), many works have applied GNNs to paper classification tasks, focusing on mining citation relationships between papers, building a paper relationship graph with papers as nodes, and applying graph learning algorithms. However, existing graph learning methods for paper classification (such as GCN and GAT) typically require labeled information about paper categories to train the model, which incurs additional data labeling costs. In contrast, graph self-supervised methods in GNNs can directly learn node representations from the graph structure without any label information. Graph contrastive learning (GCL) is such a graph self-supervised learning method applicable to GNNs. This method learns efficient representations of nodes or graphs by constructing positive and negative sample pairs and maximizing the similarity of positive pairs while minimizing the similarity of negative pairs. In paper classification tasks, the value of GCL lies in its ability to capture the complex structure and feature information of nodes (i.e., papers) in the paper relationship graph, thereby providing support for subsequent paper classification tasks.

[0004] Despite some progress in existing research, GCL still faces several challenges in the practical application of graph data augmentation for paper classification. First, traditional GCL implementations typically use contrastive loss, where positive sample pairs for each anchor are formed by generating different augmented views for the same node, while neighboring nodes are treated as negative samples to be distinguished from the anchor. However, when processing graph data, GNNs utilize neighbor aggregation mechanisms, which assume that connected nodes tend to have similar features or labels (i.e., the homogeneity assumption). Contrastive loss aims to distinguish different instances, which conflicts with the homogeneity assumption of GNNs. The second challenge is how to avoid losing crucial structural information during graph data augmentation. Common data augmentation methods include node removal and edge perturbation, which create diverse views for the model, enabling it to learn graph features from different perspectives. However, the use of these augmentation strategies can lose key structural information of the graph, leading to a mismatch between the feature representation learned by the model and the actual graph structure and function, resulting in poor performance and low accuracy in downstream paper classification tasks. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a paper classification method based on graph matching and self-supervised graph learning. The method trains on a bibliographic relationship graph structure to obtain paper classifications. First, it represents bibliographic data using a bibliographic relationship graph and proposes a novel graph learning model (ConGM) for this graph. This model leverages the powerful representational capabilities of graph neural networks to deeply mine citations and topical connections between documents, thereby achieving a more accurate classification of the fields to which papers belong.

[0006] This invention demonstrates superior performance in scientific paper classification tasks, providing a novel and effective solution for the field of document classification and retrieval. For example, a paper on semantic perception in the field of computer vision can be classified into categories such as "computer vision - classification," "computer vision - detection," "computer vision - recognition," "computer vision - segmentation," and "computer vision - object tracking" using this invention. This invention can be used to solve the problem of automatic classification of scientific papers, patent documents, images, etc., that have related or linked relationships.

[0007] Specifically, in the bibliographic graph of this invention, nodes represent papers, and edges between nodes represent citation relationships between papers. The graph learning model constructed in this invention mainly includes a linear node matching module, a quadratic edge alignment module, and a two-layer negative sample selection module to learn a comprehensive graph representation. The linear node matching module generates multiple perturbation views for the same node to construct positive sample pairs for contrastive learning, thereby encouraging the model to capture the essential attributes of nodes. The quadratic edge alignment module enhances the similarity between connected nodes, thus resolving the conflict between the homogeneity assumption and contrastive loss in graph neural networks. Furthermore, the quadratic edge alignment mechanism operates on the original bibliographic graph structure, thus preserving key structural information that might be lost during data augmentation in the node matching process. The two-layer negative sample selection module further selects difficult negative samples to enhance the understanding and representation of complex graph structures, more accurately capturing the structural and feature similarities between connected nodes.

[0008] The technical solution of the present invention is as follows:

[0009] A paper classification method based on graph matching and self-supervised graph learning includes the following steps:

[0010] 1) Construct a bibliographic graph G based on the citation relationships between the papers to be classified. The nodes of the graph represent papers (documents), and the edges between nodes represent the citation relationships between the papers and documents.

[0011] Constructing a graph learning model ConGM for paper relationship graphs, including steps 2) to 5):

[0012] 2) Subgraph sampling and data augmentation;

[0013] Two distinct nodes (node ​​u and node v) are extracted from the original literature relationship graph G, and an initial subgraph is established using a random walk method; the initial subgraph is then used to generate multiple augmented subgraphs through data augmentation.

[0014] 3) Linear node matching module: Compares and learns the document nodes in the document relationship graph;

[0015] The enhanced subgraph generated in step 2) is processed according to the node matching loss function L in step 3. n Comparative learning is performed to obtain matching node pairs and enhanced subgraphs after node matching;

[0016] 4) Secondary edge alignment module: Construction of edge center graph and comparative learning of its edges, design of edge alignment loss function;

[0017] Following step four, construct an "edge-centered graph" based on the alignment relationships of the edges. Specifically, the node matching relationships in the enhanced subgraph after node matching are used as the nodes of the edge-centered graph, and the edge matching relationships are used as the connecting edges, thereby constructing a complete edge-centered graph based on the enhanced subgraph.

[0018] Next, a graph convolutional neural network (GCN) is used to extract node feature representations from the augmented subgraph.

[0019] 5) Design a two-layer negative sample selection module / strategy to further optimize the feature representation of paper nodes;

[0020] A two-layer negative sample selection strategy is adopted to collect positive samples and hard negative samples;

[0021] In practice, for a node in the edge-centered graph, the node and the nodes connected to it are grouped into a positive sample pair; the node and other nodes (without connections) are used as intra-graph negative sample pairs in the edge-centered graph; the most challenging negative samples are calculated based on the probability distribution to obtain cross-graph negative samples; and the positive and negative samples are used as model training samples.

[0022] 6) Training the overall model;

[0023] Executing steps 2) through 5) sequentially yields a complete training model, ConGM. Based on the loss functions in steps 3) and 4), a total loss function is designed, comprising node matching loss and edge alignment loss. Training this overall model by minimizing the total loss function results in a well-trained paper classification model, learning more accurate node feature representations for each paper node.

[0024] 7) Infer, predict, and obtain paper classification results.

[0025] The dataset of document relationships to be classified is input into a pre-trained ConGM model. The model extracts features and learns representations from the input data through graph neural network layers and contrastive learning layers, generating feature vectors for each node. Subsequently, by classifying these feature vectors, the model can determine the category of the paper to which each node belongs, making inferences on the document relationship dataset and thus obtaining the corresponding paper classification results.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] This invention provides a paper classification method based on graph matching and self-supervised graph learning. It constructs a graph model to represent document data and leverages the powerful representational capabilities of graph neural networks to deeply mine citations and topical connections between documents, thereby achieving more accurate classification. Compared to traditional methods, ConGM does not rely on predefined labels but directly learns representations of nodes (i.e., documents) from the graph structure. These representations better reflect the inherent characteristics and interrelationships of the documents. This invention demonstrates superior performance in scientific paper classification tasks, providing a novel and effective solution for the field of document classification and retrieval. Attached Figure Description

[0028] Figure 1 This is a flowchart of the paper classification method based on graph matching and self-supervised graph node learning proposed in this invention.

[0029] Figure 2 This is an example of constructing a literature relationship graph based on input literature data, as an embodiment of the present invention.

[0030] Figure 3 According to Figure 2 Using the literature relationship graph in the example as input, and performing subgraph sampling, an initial subgraph G can be obtained. u ′ and G v ′.

[0031] Figure 4 After sampling the subgraph Figure 3 The example shown is a data augmentation of the results, G u The enhancement result of ′ is G u1 G u2 G v The enhancement result of ′ is G v1 G v2 .

[0032] Figure 5 After the subgraph node matching step, with Figure 4 The results shown are an example of constructing an edge-center graph.

[0033] Figure 6 After establishing the edge-center graph, with Figure 5 The results shown are examples of positive sample and two-layer negative sample selection. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the scope of the invention is not limited in any way.

[0035] As mentioned earlier, to address the shortcomings of existing GNNs used for paper classification, this invention proposes a self-supervised paper classification method based on graph matching, such as... Figure 1 As shown. This method first constructs a bibliographic graph G based on the citation relationships between the papers to be classified. For example, Figure 2 The left side shows a citation graph G constructed from 7 papers. Graph G contains 7 nodes, each corresponding to one of the 7 papers listed on the right side of the graph. Each directed edge in G represents a citation relationship between papers. For example, if paper D points to papers B and C, it means that paper D cites papers B and C. Here, we will describe the paper classification task step by step using these 7 papers as an example.

[0036] The first step is to extract nodes u and v from the original document relationship graph G, and then construct an initial subgraph G using a random walk method. u ′ and G v ',like Figure 3 As shown. These subgraphs generate multiple augmented subgraphs through data augmentation, such as G. u1 G u2 and G v1 G v2 ,like Figure 4 As shown. These enhanced subgraphs are used to capture more features of the bibliographic relationship graph.

[0037] Next, the linear node matching module performs comparative learning according to the loss function in step three, calculating the correspondence of each node in different augmented subgraphs. Then, following step four, it constructs an "edge-center graph" based on the edge alignment relationships, as shown below. Figure 5 As shown. In Figure 5 In the middle, G u1 With G u2 This is the enhanced subgraph after node matching in the previous step. There is a matching between them as follows: G u1 Nodes 1, 2, and 3 in G are respectively connected to G u2 Matching nodes a, b, and c in G u1 Edges (1,2) and (1,3) in G are respectively connected to G u2 The edges (a,b) and (a,c) in the graph are matched. Therefore, in the construction of the edge-centered graph, the nodes (1a,2b,3c) in the enhanced subgraph are used as the nodes in the edge-centered graph, and the edges are used as the edges ((1a,2b),(1a,3c)), resulting in three nodes and two edges in the edge-centered graph. Similarly, G... u1 With G u2 By constructing an edge-center graph using the above rules for all existing matches, we can obtain... Figure 5 G shown u That is, the enhanced subgraph G u1 With G u2 The complete edge-center graph is constructed. Then, a graph convolutional neural network (GCN) is used to extract the node feature representations.

[0038] Finally, a two-layer negative sample selection strategy was employed to collect suitable positive and difficult negative samples. Figure 6 For example, edge-centered graph G uIn the graph, for node 1a, its positive and negative samples are selected to form positive and negative sample pairs together with 1a: 1a and 2b are connected by an edge, so (1a, 2b) is a positive sample pair; while 1a and 1b, 1c, ..., and 2a, 3a, ... are not connected by edges, so (1a, 1b), (1a, 1c) ... and (1a, 2a), (1a, 3a) ... are selected as the edge-centered graph G. u The selection of "intra-graph negative sample pairs" involves calculating the most challenging negative sample pairs based on probability distributions. Figure 6 For example, node 1a is paired with nodes 2c and 3c. This ensures the model's learning stability across different subgraphs, thereby achieving accurate classification of the papers. Finally, the selected positive and negative sample pairs are used as training samples for comparative learning, and trained according to the loss function L in step six. This allows for the learning of more accurate node feature representations for each paper node. Classifying these feature representations using a neural network yields the specific classification results for the papers. For the seven papers in this example, the classification results are as follows:

[0039] Papers [A] and [B] are categorized under "Natural Language Processing - Word Representation," which focuses on how to better represent words or phrases in text; papers [D] and [F] are categorized under "Natural Language Processing - Pre-trained Language Models," which focuses on understanding and generating language through pre-trained deep learning models; paper [C] is categorized under "Natural Language Processing - Multi-task Learning Platforms," ​​which focuses on providing benchmarking and analysis platforms for evaluating and understanding the performance of natural language processing models across multiple tasks; and papers [E] and [G] are categorized under "Natural Language Processing - Document Processing," which focuses on handling tasks at the entire document level, such as document summarization and document classification.

[0040] The specific implementation of the method of the present invention includes the following steps:

[0041] Step 1: Construct a literature relationship diagram

[0042] This invention first constructs a bibliographic relationship graph G using paper relationship data. The construction method is to use each paper as a node to obtain a node set V, and to construct edges between nodes based on the citation relationships between papers to obtain an edge set E. The feature information of the nodes is represented by a feature matrix X, thus obtaining the bibliographic relationship graph G = (V, E, X).

[0043] Step 2: Subgraph Sampling and Data Augmentation

[0044] Select a node u in the literature relationship graph G as the reference node, and perform a random walk on the literature relationship graph G with u as the starting point and a predetermined step size s to build a subgraph (i.e., sample the subgraph of G) to generate the initial subgraph G. u The corresponding node feature matrix is ​​X'. u In a random walk, at each step, a neighboring node of the current node is randomly selected as the next node to be visited, until a predetermined step size is reached, or the subgraph G is reached. u The number of nodes N in ' u The predetermined number has been reached. In this process, the setting of parameters such as step size s determines the size and diversity of subgraphs, and s is usually best set to 10-20.

[0045] Furthermore, to generate diverse views for subsequent comparative learning, this invention performs data augmentation on the sampled subgraphs. Common data augmentation methods include node deletion, edge perturbation, and feature perturbation. Node deletion involves randomly deleting some nodes and their associated edges from the subgraph; edge perturbation involves randomly deleting or adding edges to the subgraph to disrupt the graph structure; and feature perturbation involves randomly perturbing the features of nodes, such as adding noise. Through these operations, diverse augmented subgraph views are generated to construct positive and negative sample pairs and train the model for self-supervised learning. This invention maintains a constant number of nodes on the initial subgraph G. u Edges are dropped with a predetermined probability, generating two enhanced subgraph views G. u1 G u2 The number of nodes are N1 and N2 respectively, and N1 = N2 = N u The views of edge perturbations and feature perturbations maintain consistency in the feature representation of nodes, thus enabling the capture of the essential properties of nodes.

[0046] Step 3: Linear Node Matching Module – Comparative Learning of Document Nodes in the Document Relationship Graph

[0047] This invention constructs positive sample pairs by generating multiple perturbation views, and maximizes the similarity of positive sample pairs and minimizes the similarity of negative sample pairs by minimizing the contrastive learning loss of node matching, thereby learning the structure and node feature representation of the literature relationship graph. The contrastive learning loss function for node matching (node ​​matching loss) L nAs shown in formula (1):

[0048]

[0049] in, Representing two views G u1 G u2 True matching between them Let be the node-to-node similarity matrix between the two views. τ is a hyperparameter that controls the degree of smoothness. This is achieved by minimizing the loss... This approach maximizes the inconsistency of negative sample pairs while maximizing the consistency of positive sample pairs. In this way, the model can learn the correct matching relationships between nodes, thereby gaining a better understanding of the graph's structure and node characteristics.

[0050] Ultimately, through this node-contrastive learning method, the model can learn the intrinsic attributes and complex structural features of nodes in an unsupervised environment, thereby improving the accuracy and efficiency of node classification.

[0051] Step 4: Secondary edge alignment module – Construction of edge center graph and comparative learning of edge alignment, design of edge alignment loss function;

[0052] The key to this step lies in the secondary edge alignment mechanism, which, by operating on the original literature relationship graph structure, preserves crucial structural information that might be lost during node matching due to data augmentation. The edge alignment mechanism not only focuses on the consistency of node features but also resolves the conflict between the homogeneity assumption and the contrastive loss in graph contrastive learning problems by aligning the edges connecting nodes.

[0053] In practice, different enhanced views G of node u will be used. u1 G u2 The node feature matrix X u The refined features of node u are input into the GCN model to obtain the learned node features H. u The elements in the matrix are Subsequently, construct the edge-center graph G. u These node features are mapped to a new edge-centered graph. In this graph, node V u =V u1 ×V u2 Subgraph G u1 Nodes in With subgraph G u2 Nodes in The correspondence between them, that is side This indicates the consistency between two pairs of correspondences, that is...

[0054] By calculating the edge center graph Gu The edge similarity in G is used to generate positive and negative sample pairs for contrastive learning: for G u1 Anchor nodes in and G u2 Anchor nodes in If G u1 The edge in With G u1 The edge in If we match, then we will use the edge-centered graph G. u Nodes in and nodes Considered as positive sample pairs Meanwhile, all other node pairs are considered nodes. The negative samples are paired with the negative samples to form a negative sample pair. Based on this positive and negative sample pair, the contrastive learning loss function (edge ​​alignment loss) for the edge is designed as shown in formula (2):

[0055]

[0056] Minimizing the contrastive loss function maximizes the similarity of positive sample pairs and minimizes the similarity of negative sample pairs. Here, |V| represents the total number of nodes in the original document relationship graph (i.e., the number of papers); V + V - Let represent the set of positive sample pairs and the set of negative sample pairs constructed using the method described above, respectively; h represents the node feature, and s represents the cosine similarity calculation. These are the node features learned in this step. In equation (2), the numerator represents the sum of similarities between positive sample pairs, which calculates the sum of cosine similarities between node 1a and all positive sample nodes (i.e., correctly matched nodes); similarly, the denominator calculates the sum of the similarities of the numerator plus the similarities of negative sample pairs. The entire score is then used to calculate the weight of the similarity of positive sample pairs relative to the similarity of all sample pairs (including both positive and negative sample pairs). By minimizing the loss function... This maximizes the weight, allowing the model to gradually increase the similarity of positive sample pairs and decrease the similarity of negative sample pairs during training. This learns the correct edge alignment, ensuring a more accurate graph structure representation. Step 5: Two-layer negative sample selection module—Design a two-layer negative sample pair selection strategy to further optimize the feature representation of paper nodes.

[0057] This invention introduces a two-layer negative sample selection strategy. For each node, it selects difficult negative samples both within and across the graph to form difficult negative sample pairs, further improving the model's understanding and representation of complex graph structures. Difficult negative samples are those negative samples that are very similar to positive samples, making them difficult for the model to classify correctly. Training with difficult negative samples improves the model's discriminative ability because it forces the model to learn more complex features to distinguish between positive samples and these difficult-to-distinguish negative samples.

[0058] In this approach, negative samples within the graph for each node are selected from the same graph as the current node, choosing mismatched node pairs as negative samples to form intra-graph negative sample pairs. Negative samples across graphs are selected from different graphs, using a probability distribution function to choose the most challenging negative sample pairs. Through this method, contrastive learning of edges not only optimizes the feature representation of node connectivity relationships but also effectively addresses some challenges in traditional contrastive learning.

[0059] First, as described in step four, hard negative sample pairs are identified and selected within the same graph according to graph matching rules. When constructing positive sample pairs, misaligned node pairs are treated as negative sample pairs. This process ensures that negative sample pairs are selected within the same graph, maintaining the similarity between negative and positive sample pairs. For example, for a positive sample pair... All nodes Mismatched node pairs All can be considered as nodes The negative samples are used to ensure their effective distribution in the feature space, thereby enhancing the model's discriminative ability.

[0060] In the selection of positive and negative sample pairs within a graph, for the same node, since it shares the same distribution with its positive samples, the difficult negative samples selected for it in the graph may be incorrectly selected as false negative samples. To address this issue, this invention also selects the most challenging negative sample pairs between different graphs to further improve the robustness of the model. Specifically, in the original graph G, another node v is selected as the center node, and an initial subgraph G is generated in the same manner as the anchor node u. v ′ and enhanced subgraph G v1 G v2 , based on G, through the calculation of probability distribution. v1 G v2 Construct a new edge-center graph G v Calculate the probability distribution function f of the node features in the two edge centroid graphs. s∣u (s) and f s∣v (s), where s is any node. By comparing two probability distributions, if a node pair has a probability exceeding a certain threshold T in one graph, but is a negative sample of the current node in the other graph, then it is selected as the hard negative sample of the current node.

[0061] This invention employs a two-layer negative sample selection strategy, which not only ensures the effectiveness and similarity of negative samples but also enhances the model's robustness and discriminative ability by introducing negative samples from different graphs. This method enables the model to better capture subtle differences in complex graph structures, significantly improving its performance in handling complex graph structures and performing self-supervised learning tasks.

[0062] Step Six: Training the Overall ConGM Model

[0063] Combining steps two through four, a complete model can be obtained. This model includes the subgraph sampling and data augmentation steps, linear node matching module, quadratic edge alignment module, and two-layer negative sample sampling module described above. Among these, the node matching loss in the linear node matching module... The primary focus is on maximizing the consistency of positive sample pairs, while the edge alignment loss in the secondary edge alignment module... This approach aims to enhance the similarity between connected nodes, thereby resolving the conflict between the homogeneity assumption and contrastive loss in GNNs. Since the model needs to optimize both loss functions simultaneously during training, a node matching loss mechanism must be incorporated. and edge alignment loss The overall loss L is obtained:

[0064] L = L n +βL e (3)

[0065] The above formula balances the two losses using an appropriate loss coefficient β (β is optimally around 0.1). This allows the model to effectively learn node and graph representations in graph self-supervised learning, while simultaneously addressing the challenges posed by the homogeneity assumption and the potential loss of structural information due to data augmentation. By training the input literature relationship graph using the aforementioned loss function L, the ConGM model can be obtained.

[0066] Step 7: Infer, predict, and obtain the paper classification results

[0067] The dataset of document relationships to be classified is input into the pre-trained ConGM model. The model extracts features and learns representations from the input data through the Graph Neural Network (GCN) layer and the contrastive learning layers in steps four and five, thus obtaining the vector representations of the nodes. Subsequently, by classifying these node vectors through a single-layer neural network, the paper category to which each node belongs can be obtained. Inferences are made on the document relationship dataset to obtain the corresponding paper classification results.

[0068] It should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the scope of the claims.

Claims

1. A paper classification method based on graph matching and self-supervised graph learning, characterized in that, First, the literature data is represented by a bibliographic relationship graph. Then, a graph learning model ConGM based on this graph is constructed, including subgraph sampling and data augmentation modules, linear node matching modules, quadratic edge alignment modules, and two-layer negative sample selection modules. This model mines citations and topical connections between documents, thereby achieving accurate classification of the papers' domains. The steps include: 1) Construct a bibliographic graph based on the citation relationships between the papers to be classified; the nodes of the graph represent documents, and the edges between nodes represent the citation relationships between the documents. Constructing a graph learning model ConGM based on document relationship graphs, i.e., a paper classification model; including steps 2) to 5): 2) Subgraph sampling and data augmentation: Two different nodes are extracted from the literature relationship graph, and an initial subgraph is established by random walk method; Then, based on the initial subgraph, multiple enhanced subgraph views are generated through data augmentation; 3) Linear node matching: Comparative learning is performed on the document nodes of the document relationship graph to obtain matching node pairs and enhanced subgraphs after node matching; Specifically, positive sample pairs are constructed by generating multiple perturbations, and the similarity of positive sample pairs is maximized and the similarity of negative sample pairs is minimized by using structured linear loss and contrastive learning loss, i.e., minimizing contrastive learning loss, thereby learning the structure and node feature representation of the literature relationship graph. 4) Based on the alignment relationship of the edges, the node matching relationship in the enhanced subgraph after node matching is used as the node of the edge center graph, and the edge matching relationship is used as the connecting edge. The edge center graph and its edge contrast learning are constructed; the edge contrast loss function is designed; and the node feature representation in the subgraph is extracted using a graph convolutional neural network. 5) Design a two-layer negative sample selection strategy to further optimize the feature representation of document nodes and collect positive samples and hard negative samples; The specific process is as follows: For each node in the edge-centered graph, the node and the nodes connected to it are grouped into a positive sample pair; the node and other nodes without connections are grouped into negative sample pairs within the edge-centered graph; the most challenging negative sample is calculated based on the probability distribution to obtain cross-graph negative samples; and the positive and negative samples are used as model training samples. 6) Perform model training; Design a total loss function, which includes node matching loss and edge alignment loss. Train the model to obtain a well-trained paper classification model, which learns a more accurate node feature representation for each paper node. 7) Using the trained paper classification model, perform inference and prediction to obtain the paper classification results.

2. The paper classification method based on graph matching and self-supervised graph learning as described in claim 1, characterized in that, In step 2), the data augmentation methods include node deletion, edge perturbation, and feature perturbation.

3. The paper classification method based on graph matching and self-supervised graph learning as described in claim 1, characterized in that, In step 3), the contrastive loss function L for node matching n Represented as Equation (1): in, Represents two enhanced views G u1 G u2 True matching between them Let be the node-to-node similarity matrix between the two views; τ is a hyperparameter that controls the degree of smoothness; by minimizing the contrast loss, the inconsistency of negative sample pairs is maximized while the consistency of positive sample pairs is maximized, thereby learning the correct matching relationship between nodes.

4. The paper classification method based on graph matching and self-supervised graph learning as described in claim 3, characterized in that, In step 4), construct the edge-center graph G. u Map node features to an edge-centered graph; nodes V in the edge-centered graph u =V u1 ×V u2 Represents the enhanced subgraph G u1 Nodes in With the enhancement subgraph G u2 Nodes in The correspondence between them, that is side This indicates the consistency between two pairs of correspondences, that is...

5. The paper classification method based on graph matching and self-supervised graph learning as described in claim 4, characterized in that, In step 4), the edge comparison loss function is used. Represented as: Where h represents node features, and s represents the cosine similarity calculation. The learned node features; |V| represents the total number of nodes in the document relationship graph, i.e., the number of papers; V + V - Let them represent the set of positive sample pairs and the set of negative sample pairs, respectively.

6. The paper classification method based on graph matching and self-supervised graph learning as described in claim 5, characterized in that, In step 5), the two-layer negative sample selection strategy identifies and selects hard negative sample pairs within the same graph based on graph matching methods; specifically, it includes: When constructing positive sample pairs, misaligned node pairs are treated as negative sample pairs; In selecting positive and negative sample pairs within a graph, the most challenging negative sample pair is chosen from different graphs; the specific steps are as follows: In the original graph G, select another node v as the center node, and generate the initial subgraph G in the same way as the anchor node u. v ′ and enhanced subgraph G v1 G v2 ; Based on G v1 G v2 Construct a new edge-center graph G v Calculate the probability distribution function f of the node features in the two edge centroid graphs. s∣u (s) and f s∣v (s), where s is any node; By comparing two probability distributions, if a node pair has a probability exceeding a certain threshold T in one graph and is a negative sample of the current node in the other graph, then it is selected as a hard negative sample of the current node.

7. The paper classification method based on graph matching and self-supervised graph learning as described in claim 5, characterized in that, In step 6), the node matching loss L n To maximize the consistency of positive sample pairs, the edge alignment loss L e Used to enhance the similarity between connected nodes; During training, the node matching loss L is optimized. n And edge alignment loss L e The overall loss L is obtained and expressed as: L=L n +βL e , (3) Where β is the loss coefficient that balances the two losses.

8. The paper classification method based on graph matching and self-supervised graph learning as described in claim 7, characterized in that, In step 6), the loss coefficient is set to 0.1.

Images