Graph structure data classification method based on dynamic multi-factor decoupling and related equipment

Abstract

The embodiment of the invention relates to the technical field of machine learning, and discloses a graph structure data classification method and related equipment based on dynamic multi-factor decoupling, and the method comprises the steps: obtaining graph data and an undirected graph constructed according to the graph data; inputting the graph data into a graph decoupling network, and executing coarse-grained dynamic multi-factor decoupling and fine-grained double constraint so as to perform coarse-to-fine hierarchical decoupling; determining a factor-level prototype contrast loss based on the embedded update representation; obtaining a category prediction result of the node based on a preset classifier, and determining cross entropy loss based on the category prediction result and the category label of the node; constructing an overall optimization target in combination with cross entropy loss, geometric diversity loss and factor-level prototype comparison loss; and optimizing parameters of the graph decoupling network based on the overall optimization target so as to perform node classification based on the optimized graph decoupling network. By means of the mode, the category separability can be improved, and therefore the node classification performance is improved.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, specifically to a graph structure data classification method and related equipment based on dynamic multi-factor decoupling. Background Technology

[0002] Node classification is a core task in learning graph-structured data, aiming to predict node categories by combining node features with neighborhood structural relationships within the graph. It has wide applications in social network analysis, recommender systems, knowledge graph reasoning, and molecular attribute prediction. Graph neural networks, as deep learning models for modeling graph-structured data, learn richer feature representations by aggregating node neighbor information, demonstrating excellence in tasks such as node classification, graph classification, and link prediction, and have become the mainstream method for node classification. However, most graph neural networks typically treat the node's neighborhood as a whole for feature aggregation during message passing, ignoring the subtle differences caused by different semantic relationships or generative factors within the neighborhood. In fact, graph structures in the real world are often formed by the interaction of multiple underlying factors. If neighborhood information is aggregated indiscriminately, the model will struggle to explicitly identify and model these underlying factors, resulting in node representations with mixed semantics and a lack of decoupling, thus limiting interpretability and discriminability.

[0003] Graph decoupling representation learning can alleviate the aforementioned problems. It is an interdisciplinary field combining decoupling representation learning and graph neural networks. Its core idea is to decouple multiple latent factors in graph-structured data and learn semantically meaningful decoupling representations. This overcomes the interpretability and discriminative limitations of traditional graph neural networks, which employ holistic aggregation strategies and ignore differences within neighborhoods. Currently, there are various graph decoupling representation learning methods, but existing ones generally focus on improving intra-factor consistency and inter-factor independence, neglecting the class separability of the decoupled latent factor space. This limits downstream classification performance, which still needs improvement. Summary of the Invention

[0004] In view of the above problems, embodiments of the present invention provide a graph structure data classification method and related equipment based on dynamic multi-factor decoupling, which are used to solve the problems existing in the prior art.

[0005] According to one aspect of the present invention, a graph structure data classification method based on dynamic multi-factor decoupling is provided, the method comprising: Acquire graph data and an undirected graph constructed based on the graph data. The undirected graph includes a set of nodes and an edge set consisting of multiple nodes. Each node is composed of multiple independent latent factors, and each latent factor is assigned a latent factor space. The graph data is input into a graph decoupling network, and coarse-grained dynamic multi-factor decoupling and fine-grained dual constraints are performed to decouple multiple potential factors of the node in a hierarchical manner from coarse to fine. The fine-grained dual constraints include applying structural constraints and geometric constraints to the potential factor space based on the coarse-grained decoupling representation to decouple at the fine-grained level. After applying the structural constraints, the embedding update representation of the node in the potential factor space is obtained. The geometric constraints are imposed through geometric diversity loss. The factor-level prototype contrast loss is determined based on the embedded update representation to enhance factor-level intra-class aggregation and inter-class separation; The class prediction result of the node is obtained based on a preset classifier, and the cross-entropy loss is determined based on the class prediction result and the class label of the node. The overall optimization objective is constructed by combining the cross-entropy loss, the geometric diversity loss, and the factor-level prototype contrast loss. The parameters of the graph decoupled network are optimized based on the overall optimization objective, so as to perform node classification based on the optimized graph decoupled network.

[0006] In one alternative approach, the fine-grained dual constraint includes applying structural and geometric constraints to the latent factor space based on the coarse-grained decoupling representation, respectively, to achieve decoupling at the fine-grained level, including: For each of the latent factor spaces, a corresponding K-nearest neighbor graph is constructed. The construction of the corresponding K-nearest neighbor graph includes: determining the Euclidean distance between the nodes based on the coarse-grained decoupled representation; determining the adjacency matrix of the nodes based on the Euclidean distance; constructing the K-nearest neighbor graph based on the coarse-grained decoupled representation and the adjacency matrix; applying a graph convolutional network to the K-nearest neighbor graph to perform an aggregation operation, obtaining the embedding update representation of the nodes in the latent factor space; and concatenating the embedding update representations of all latent factor spaces to obtain a compact decoupled representation. Based on the coarse-grained decoupling representation, the statistical response vectors of each node in different potential factor spaces are extracted. The statistical response vectors are normalized to obtain a normalized response matrix. The subspace volume spanned by the statistical response vectors is calculated based on the normalized response matrix. A geometric diversity regularization term is constructed based on the subspace volume, and the regularization term is used as the geometric diversity loss.

[0007] In one alternative approach, determining the factor-level prototype contrastive loss based on the embedded updated representation includes: Based on the embedding update representation of similar nodes in the training set, a category prototype is constructed. ; , in, Let be the class prototype of class c in the m-th potential factor space. The set of nodes belonging to class c in the training set. Let u be the embedding update representation of the node in the m-th potential factor space; Calculate the normalized cosine similarity between the node and the category prototype; The prototype contrast loss is determined based on the normalized cosine similarity. : , in, The set of nodes in the training set. For the normalized cosine similarity, For the node Category tags, Temperature coefficient; Prototype contrast loss across all potential factor spaces Perform a joint summation to obtain the factor-level prototype contrast loss.

[0008] In one optional approach, the coarse-grained dynamic multi-factor decoupling includes: mapping the node features to multiple potential factor spaces, calculating hub scores and authority scores based on the similarity between the node and its neighboring nodes in different potential factor spaces, determining aggregated representations of each potential factor space based on the hub scores, the authority scores, and dynamic differential weights, and concatenating the aggregated representations of all potential factor spaces to obtain the coarse-grained decoupling representation.

[0009] In one alternative approach, constructing the overall optimization objective by combining the cross-entropy loss, the geometric diversity loss, and the factor-level prototype contrast loss includes constructing the overall optimization objective L using the following formula: , in, For the cross-entropy loss, For the geometric diversity loss, This refers to the factor-level prototype contrast loss. Hyperparameters are used to balance the weights of various losses.

[0010] In an alternative approach, the method further includes: adjusting the hyperparameters on multiple benchmark datasets. Sensitivity analysis was performed.

[0011] In an alternative approach, the method further includes: Visualization operations are performed during node classification based on the optimized graph decoupled network. These visualization operations include visualization of node embeddings, visualization of node pair similarity in different latent factor spaces, visualization of latent factor decoupling, visualization of latent feature correlation, and visualization of factor-level prototype contrastive loss embeddings.

[0012] According to another aspect of the present invention, a graph structure data classification device based on dynamic multi-factor decoupling is provided, the device comprising: The acquisition module is used to acquire graph data and an undirected graph constructed based on the graph data. The undirected graph includes a node set and an edge set consisting of multiple nodes. Each node is composed of multiple independent latent factors, and each latent factor is assigned a latent factor space. The hierarchical decoupling module is used to input the graph data into the graph decoupling network, perform coarse-grained dynamic multi-factor decoupling and fine-grained dual constraints, so as to perform hierarchical decoupling of multiple potential factors of the node from coarse to fine. The fine-grained dual constraints include applying structural constraints and geometric constraints to the potential factor space based on the coarse-grained decoupling representation, so as to decouple at the fine-grained level. After applying the structural constraints, the embedding update representation of the node in the potential factor space is obtained. The geometric constraints are constrained by geometric diversity loss. The first determining module is used to determine the factor-level prototype contrast loss based on the embedded update representation, so as to enhance factor-level intra-class aggregation and inter-class separation; The second determining module is used to obtain the category prediction result of the node based on a preset classifier, and to determine the cross-entropy loss based on the category prediction result and the category label of the node. The construction module is used to combine the cross-entropy loss, the geometric diversity loss, and the factor-level prototype contrast loss to construct an overall optimization objective; The classification module is used to optimize the parameters of the graph decoupled network based on the overall optimization objective, so as to classify nodes based on the optimized graph decoupled network.

[0013] According to another aspect of the present invention, a computer device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; the memory is used to store at least one executable instruction, wherein the executable instruction causes the processor to perform the method described above.

[0014] According to another aspect of the present invention, a computer-readable storage medium is provided, the storage medium storing at least one executable instruction, which, when executed on a computer device, causes the computer device to perform the method described above.

[0015] This invention employs a coarse-to-fine hierarchical decoupling approach. First, a dynamic allocation mechanism is used to achieve coarse-grained dynamic decoupling of latent factors. Then, fine-grained structural and geometric constraints are applied to achieve progressive and refined optimization of decoupling quality. This effectively enhances the consistency within latent factors and the diversity between latent factors, ensuring the high quality of the final decoupled representation. Finally, a coordinating factor-level prototype contrast loss is used to improve intra-class aggregation and inter-class separation in the decoupled latent factor spaces, thereby improving class separability and significantly enhancing node classification performance.

[0016] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0017] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 The flowchart shown is an embodiment of the graph structure data classification method based on dynamic multi-factor decoupling provided by the present invention. Figure 2 A schematic diagram of a graph decoupling network according to an embodiment of the present invention is shown; Figure 3 The graph decoupling network of this invention and various baseline models are shown in the node embedding visualization results on the Cora dataset; Figure 4 The visualization results of the latent factor space node similarity matrix obtained by different models on the Cora test set are shown; Figure 5 The visualization results of different models decoupling latent factors on the Cora dataset are shown; Figure 6 The component confusion matrix visualization results for different models on the Cora dataset are shown; Figure 7 The diagram shows the distribution of nodes in each potential factor space with and without using factor-level prototype contrast loss. Figure 8 The ablation experiment results of the graph decoupling network of this invention on six benchmark datasets are shown; Figure 9 This diagram illustrates the structure of a graph-structured data classification device based on dynamic multi-factor decoupling according to an embodiment of the present invention. Figure 10A schematic diagram of the structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation

[0018] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0019] To achieve high-quality decoupling of latent factors and enhance the discriminability of each latent factor space, this invention proposes a graph-structured data classification method based on dynamic multi-factor decoupling. The graph decoupling network includes a dynamic multi-factor hierarchical decoupling module and a factor-level discriminability enhancement module. The dynamic multi-factor hierarchical decoupling module effectively achieves hierarchical decoupling of latent factors by integrating a dynamic allocation mechanism and dual structural and geometric constraints, thereby obtaining high-quality decoupling representations. Secondly, the factor-level discriminability enhancement module constructs a class prototype-based contrastive learning mechanism in each factor space and proposes a factor-level prototype-guided contrastive loss, explicitly optimizing intra-class aggregation and inter-class separation at the factor level, thus effectively improving the discriminability of each latent factor. This invention achieves significant improvements in node classification performance, latent factor decoupling quality, and class separability of each factor space, fully validating the effectiveness of the model.

[0020] Figure 1 The flowchart of an embodiment of the graph structure data classification method based on dynamic multi-factor decoupling provided by the present invention is shown, as follows: Figure 1 As shown, the method includes the following steps: Step 10: Obtain graph data and an undirected graph constructed based on the graph data. The undirected graph includes a set of nodes and an edge set consisting of multiple nodes. Each node is composed of multiple independent latent factors, and each latent factor is assigned a latent factor space. Given an undirected graph G = (V, E), where V and E are the node set and edge set, respectively, and the node set contains N nodes. The characteristic matrix of the node set is... ,in Let f represent the feature vector of node u, where f is the dimension of the input features. Suppose that node u consists of M independent latent factors, i.e. Let d represent the embedding vector of node u in the m-th latent factor space, and d be the representation dimension of each latent factor space.

[0021] To characterize the connectivity between nodes, an adjacency matrix is ​​defined. ,in, This indicates that there is an edge between node u and node v. For any node... Its set of neighboring nodes is denoted as .

[0022] Step 20: Input the graph data into the graph decoupling network and perform coarse-grained dynamic multi-factor decoupling and fine-grained dual constraints to decouple multiple potential factors of the node in a hierarchical manner from coarse to fine. The fine-grained dual constraints include applying structural constraints and geometric constraints to the potential factor space based on the coarse-grained decoupling representation to decouple at the fine-grained level. After applying the structural constraints, the embedding update representation of the node in the potential factor space is obtained. The geometric constraints are imposed through geometric diversity loss. To achieve high-quality decoupled node representation, the graph decoupling network in this embodiment adopts a coarse-to-fine hierarchical decoupling strategy: at the coarse-grained level, a dynamic allocation mechanism is used to flexibly model the multiple relationships between nodes and their neighbors to achieve decoupling of potential factors; at the fine-grained level, constraints are imposed on the subspace of each factor from both structural and geometric perspectives to optimize the spatial distribution of potential factors, thereby enhancing consistency within factors and diversity between factors.

[0023] Furthermore, the coarse-grained dynamic multi-factor decoupling includes: mapping the node features to multiple potential factor spaces, calculating hub scores and authority scores based on the similarity between the node and its neighboring nodes in different potential factor spaces, determining the aggregated representation of each potential factor space based on the hub scores, the authority scores, and dynamic differential weights, and concatenating the aggregated representations of all potential factor spaces to obtain the coarse-grained decoupling representation.

[0024] To obtain a coarse-grained decoupled representation of nodes, this embodiment employs a dynamic allocation mechanism to decouple latent factors. First, the feature vectors of nodes are mapped to M latent factor spaces, thereby initializing the embedding representation of each latent factor space. , , in, Let these represent the weight matrix and bias term of the m-th potential factor space, respectively. It is a non-linear activation function.

[0025] Next, based on the similarity between node u and its neighbor node v in the m-th latent factor space, the hub score and authority score are calculated respectively, where the hub score of node u in the m-th factor space is... And authority score The calculation is as follows: ; ; in, Let represent the similarity between node u and its neighbor node v in the m-th latent factor space. It is the set of neighboring nodes.

[0026] Neighboring nodes participate in the aggregated representation of multiple latent factor spaces with dynamic differential weights. The aggregated representation of node u in the m-th latent factor space is updated according to the following formula: , in, This represents the aggregation result obtained in the previous iteration. There are two trainable parameters.

[0027] Through the above steps, the embedding representation of the m-th latent factor space can be obtained. By concatenating the embedding representations of all M latent factor spaces, we can obtain the coarse-grained decoupled representation of the nodes. .

[0028] Furthermore, the fine-grained dual constraint includes applying structural and geometric constraints to the latent factor space based on the coarse-grained decoupling representation, respectively, to achieve decoupling at the fine-grained level, including: For each of the latent factor spaces, a corresponding K-nearest neighbor graph is constructed. The construction of the corresponding K-nearest neighbor graph includes: determining the Euclidean distance between the nodes based on the coarse-grained decoupled representation; determining the adjacency matrix of the nodes based on the Euclidean distance; constructing the K-nearest neighbor graph based on the coarse-grained decoupled representation and the adjacency matrix; applying a graph convolutional network to the K-nearest neighbor graph to perform an aggregation operation, obtaining the embedding update representation of the nodes in the latent factor space; and concatenating the embedding update representations of all latent factor spaces to obtain a compact decoupled representation. Based on the coarse-grained decoupling representation, the statistical response vectors of each node in different potential factor spaces are extracted. The statistical response vectors are normalized to obtain a normalized response matrix. The subspace volume spanned by the statistical response vectors is calculated based on the normalized response matrix. A geometric diversity regularization term is constructed based on the subspace volume, and the regularization term is used as the geometric diversity loss.

[0029] After obtaining the coarse-grained decoupled representation, the feature distribution of nodes in different latent factor spaces still has shortcomings. On the one hand, the node distribution within the same latent factor space is too loose and lacks local consistency. On the other hand, there may be representational redundancy or semantic overlap between different latent factor spaces, weakening the decoupling effect. To address this, this embodiment imposes constraints on the latent factor space from both structural and geometric perspectives, strengthening the consistency within latent factors and the diversity between latent factors respectively, thereby achieving higher-quality decoupled representation at the fine-grained level.

[0030] To enhance consistency within the latent factors, this embodiment constructs a specific K-nearest neighbor graph G for each latent factor space m, based on the coarse-grained decoupled representation. m This is to characterize the local structural dependencies of each latent factor space. First, based on the embedding representation Z of the m-th latent factor space... (m) Calculate the Euclidean distance between nodes, and obtain the adjacency matrix A based on the Euclidean distance. (m) : , in, Let represent the set of k nearest neighbors of node u in the m-th potential factor space by Euclidean distance.

[0031] Based on adjacency matrix A (m) Construct a K-nearest neighbor graph (G1, G2, ..., G...) M ),in, Next, for each K-nearest neighbor graph G m An aggregation operation is performed using a graph convolutional network to update the embedding representation of a node in the m-th latent factor space, thus obtaining the updated embedding representation. Finally, the embedding update representation of all latent factor spaces. By splicing, a more compact decoupled representation within the latent factors is obtained. This achieves structural constraints.

[0032] To enhance the diversity among latent factors, this embodiment imposes constraints on each latent factor space from a geometric perspective. The directional dispersion is characterized by calculating the volume of the subspace spanned by each latent factor vector, thereby encouraging good discriminability among latent factors in the representation space. First, based on the embedding vector z of node u in the m-th latent factor space... u,m Extract its corresponding statistical response vector R u,m Next, the statistical response vectors of each potential factor are normalized to obtain the normalized response matrix. This eliminates scale differences across different potential factor spaces while preserving directional information. Finally, the normalized response matrix is ​​calculated. The Gram determinant is used to obtain the volume of the subspace spanned by the statistical response vectors of each latent factor. : , Among them, the volume of Zhang Cheng's subspace A measure of the geometric diversity of node u. This represents the calculation of determinants. It represents the volume of the subspace spanned when vectors are orthogonal or nearly independent. A larger value indicates greater diversity; conversely, a smaller value indicates redundancy. This is based on the volume of the spanned subspace. Construct a geometric diversity regularization term and use the regularization term as the geometric diversity loss. : ,

[0033] Through geometric diversity loss Implement geometric constraints to enhance the diversity among potential factors.

[0034] Step 30: Determine the factor-level prototype contrast loss based on the embedded update representation to enhance factor-level intra-class aggregation and inter-class separation; Existing graph decoupling techniques (such as DisenGCN and ADGCN) primarily focus on separating factors (independence between factors), but neglect the discriminative power of each factor space after separation. This embodiment, through factor-level prototypical contrastive loss, explicitly takes improving discriminative power as a core optimization objective of the decoupling process, achieving a paradigm shift from simple decoupling to a synergistic enhancement of decoupling and discriminative power.

[0035] This embodiment achieves high-quality decoupled node representations after performing coarse-to-fine hierarchical decoupling. However, even after decoupling, intra-class dispersion and inter-class overlap may still occur in each latent factor space, limiting the performance of the decoupled node representations in node classification tasks. Therefore, the graph decoupling network in this embodiment further enhances the class separability of nodes in each latent factor space through factor-level prototype contrastive loss, thereby obtaining higher-quality, more discriminative node representations to improve node classification performance. Specifically, this includes constructing a class prototype-based contrastive mechanism in each latent factor space, causing nodes to aggregate towards the semantic center of their respective class and move away from other class centers, achieving factor-level intra-class compactness and inter-class separation.

[0036] Further, the factor-level prototype contrastive loss is determined based on the embedded update representation, including: Based on the embedding update representation of similar nodes in the training set, a category prototype is constructed. : , in, Let be the class prototype of class c in the m-th potential factor space. The set of nodes belonging to class c in the training set. This represents the embedding update of node u in the m-th latent factor space. This embodiment's category prototype... The statistical semantic center of each category in the m-th latent factor space is characterized. Unlike the traditional static category center, the category prototype used in this embodiment is dynamically updated with the node representation during the training process, which can adaptively characterize the constantly changing semantic structure in the decoupled space. This mechanism can effectively reduce the statistical bias caused by small-batch training and improve the stability of category center estimation.

[0037] Calculate the normalized cosine similarity between the node and the category prototype. : , Among them, normalized cosine similarity As a unified metric between nodes and category prototypes.

[0038] The prototype contrast loss is determined based on the normalized cosine similarity. : , in, The set of nodes in the training set. For the normalized cosine similarity, The category label for node u. Temperature coefficient. Prototype comparison loss. Essentially, it's a contrastive learning objective that uses category prototypes as a reference: the numerator encourages nodes to maintain higher similarity to their corresponding category prototypes, thus strengthening intra-class aggregation; the denominator, through competitive normalization with all category prototypes, implicitly widens the similarity gap between different category prototypes, thereby enhancing inter-class separation, thus achieving prototype contrast loss. It can improve the class separability of nodes in each potential factor space.

[0039] Prototype contrast loss across all potential factor spaces Perform a joint summation to obtain the factor-level prototype contrast loss.

[0040] Unlike traditional supervised comparisons that only operate in the original space, the factor-level prototype comparison loss in this embodiment uses the category prototype as a global semantic reference. Through the comparison learning mechanism between nodes and category prototypes, it effectively alleviates the noise interference and instability of traditional sample comparison methods. Furthermore, it is executed on each latent factor in the decoupled multi-latent factor space, which can work synergistically with the decoupled structure to improve the discriminativeness of each latent factor space.

[0041] Step 40: Obtain the category prediction result of the node based on the preset classifier, and determine the cross-entropy loss based on the category prediction result and the category label of the node; The classifier can be an existing classifier, and the class prediction result for a node includes the probability that node u belongs to class c. To measure the difference between the prediction result of the graph decoupling network and the true class label, this embodiment calculates the cross-entropy loss based on the class prediction result and the class label of the node. : , in, Let C be the set of nodes in the training set, and C be the number of classes. This is the indicator value of the actual category label of node u in category c. Let be the probability that node u belongs to category c.

[0042] Step 50: Combine the cross-entropy loss, the geometric diversity loss, and the factor-level prototype contrast loss to construct an overall optimization objective; The overall optimization objective L is constructed using the following formula: , in, For the cross-entropy loss, For the geometric diversity loss, This refers to the factor-level prototype contrast loss. Hyperparameters are used to balance the weights of various losses.

[0043] Cross-entropy loss It can measure the difference between the prediction results of graph decoupling networks and the true category labels; geometric diversity loss It can suppress feature redundancy and geometric diversity loss among potential factors. The diversity among latent factors is characterized by calculating the volume of the subspace spanned by the statistical response vectors corresponding to different latent factor spaces, thereby enabling the graph decoupling network to learn complementary and non-overlapping latent factor representations; factor-level prototype contrastive loss. To improve the separability of categories in each latent factor space, it prompts nodes to aggregate toward their category prototypes in each latent factor space and distance themselves from other category prototypes through supervision signals guided by category prototypes, thereby explicitly strengthening intra-class aggregation and inter-class separation.

[0044] Step 60: Optimize the parameters of the graph decoupling network based on the overall optimization objective, and classify nodes based on the optimized graph decoupling network.

[0045] This embodiment optimizes the parameters of the graph decoupling network by optimizing the overall optimization objective L. Specifically, the values ​​of all learnable parameters can be adjusted using the gradient descent algorithm to improve the classification performance of the graph decoupling network. The learnable parameters may include one or more of the following: feature mapping and initialization parameters, parameters related to the dynamic allocation mechanism, parameters of the graph convolutional network, parameters of the classifier, and parameters of the category prototype.

[0046] This invention employs a coarse-to-fine hierarchical decoupling approach. First, a dynamic allocation mechanism is used to achieve coarse-grained dynamic decoupling of latent factors. Then, fine-grained structural and geometric constraints are applied to achieve progressive and refined optimization of decoupling quality. This effectively enhances the consistency within latent factors and the diversity between latent factors, ensuring the high quality of the final decoupled representation. Finally, a coordinating factor-level prototype contrast loss is used to improve intra-class aggregation and inter-class separation in the decoupled latent factor spaces, thereby improving class separability and significantly enhancing node classification performance.

[0047] See also Figure 2 The graph decoupling network consists of a dynamic multi-factor hierarchical decoupling module a and a factor-level discrimination enhancement module b. The input A is the adjacency matrix, and X is the feature matrix. In the dynamic multi-factor hierarchical decoupling module a, geometric diversity loss is used... Geometric constraints are applied, and factor-level discrimination enhancement module b uses factor-level prototype comparison loss. The module a, which enhances intra-class aggregation and inter-class separation, achieves high-quality decoupling of potential factors through a coarse-to-fine hierarchical design. The factor-level discrimination enhancement module b can improve the class separability of each potential factor space after decoupling. The two work together to significantly improve the node classification performance.

[0048] Furthermore, embodiments of the present invention evaluate the above-described graph decoupling network on six real-world graph datasets. The six real-world graph datasets include: 1. Cora, Citeseer, and PubMed are three classic citation network datasets, primarily used for node classification tasks. In these graphs, nodes, edges, and labels represent documents, citation relationships between documents, and the research fields to which the documents belong, respectively.

[0049] 2. Computers and Photo are two classic datasets from co-purchasing networks, commonly used for product recommendation and social network analysis. Nodes represent products, edges indicate that two products are frequently purchased together, and labels represent the category to which the product belongs.

[0050] 3. CS is a co-authored network dataset from Microsoft Academic Graph. This dataset mainly describes the collaborative relationships between academic authors, where nodes represent academic authors, edges indicate that two academic authors have a co-authored relationship, that is, they have co-authored at least one paper, and labels indicate the academic field of the author.

[0051] To ensure consistency in settings across different datasets and to fairly compare the performance of each model, this embodiment of the invention employs two data partitioning strategies. For the Cora, Citeseer, and PubMed datasets, the standard partitioning method in Graph Convolutional Networks (GCNs) is used: 20 nodes are selected for each category as the training set, 500 nodes as the validation set, and 1000 nodes as the test set. For the Computers, Photo, and CS datasets, a random partitioning method is used: 20 nodes are selected for each category for the training set, 30 nodes for each category for the validation set, and the remaining nodes as the test set.

[0052] To verify the effectiveness of the graph decoupling network described above, a series of representative and advanced algorithms were selected for systematic comparison, covering six mainstream deep graph neural network models GCN, GAT, SGC, AP-GCN, ARMA, and RNCGLN, and six representative graph decoupling networks DisenGCN, IPGDN, FactorGCN, SGCN, LGD-GCN, and ADGCN.

[0053] The embodiments of the present invention compare performance through node classification and cluster analysis.

[0054] 1. Node Classification: Table 1 below shows a comparison of the experimental results of the above-mentioned graph decoupling network (Ours) and various baseline models in the node classification task, with the node classification results of the embodiment of the present invention being the best. As can be seen from Table 1, the decoupling method generally outperforms traditional graph neural networks (GCN, GAT, and SGC) in performance. Furthermore, the graph decoupling network of the embodiment of the present invention outperforms the six existing graph decoupling methods on multiple datasets, with a performance improvement of over 1.4% compared to the best-performing baseline model on the Citeseer and Photo datasets. Specifically, the comparison results with the representative method DisenGCN show that the graph decoupling network of the embodiment of the present invention improves the accuracy by approximately 1.8%, 3.1%, 1.6%, 1%, 2%, and 1.8% on the Cora, Citeseer, Pubmed, CS, Computers, and Photo datasets, respectively. Table 1 2. Cluster analysis: To further verify the effectiveness of the graph decoupling network in node representation learning according to the embodiments of the present invention, clustering analysis was performed on the generated embeddings. Specifically, node embedding vectors were extracted from the output of the graph decoupling network and clustered using the classic KMeans algorithm, where the number of clusters K was set as the number of true categories. To comprehensively evaluate the clustering effect, three commonly used clustering performance metrics were selected: accuracy (ACC), normalized mutual information (NMI), and average Rand index (ARI), to measure the quality of node embeddings from different perspectives. Table 2 shows the comparison of clustering performance of different methods on the citation network dataset, where the clustering performance of the embodiment of the present invention is the best, significantly better than traditional graph neural networks. In addition, among various graph decoupling methods, the graph decoupling network of the embodiment of the present invention achieves the best results in all clustering metrics. This is mainly due to its ability to construct a comparison mechanism based on category prototypes in each latent factor space while achieving decoupling of latent factors, thereby strengthening intra-class aggregation and inter-class separation at the latent factor level. This factor-level discriminative enhancement effectively improves the global separability of the embedding, enabling the node representation to exhibit clearer inter-class boundaries and a more compact intra-class distribution in the semantic space, thereby resulting in better clustering performance. Table 2 Furthermore, in an optional embodiment, the method further includes performing visualization operations during node classification based on the optimized graph decoupled network, the visualization operations including visualization of node embeddings, visualization of node pair similarity in different latent factor spaces, visualization of latent factor decoupling, visualization of latent feature correlation, and visualization of factor-level prototype contrastive loss embeddings.

[0055] The embodiments of the present invention further verify the effectiveness of graph decoupling networks through a series of visualization analyses.

[0056] 1. Node Embedding Visualization: To intuitively verify the discriminative ability of the graph decoupling network in node representation learning, a visualization analysis of the node embeddings generated in the last layer was performed. Specifically, the t-SNE dimensionality reduction algorithm was used to project the high-dimensional embedding vectors into a two-dimensional space, and the nodes were labeled with different colors according to their category labels.

[0057] Figure 3This paper presents the visualization results of node embeddings of the graph decoupled network (Ours) and various baseline models on the Cora dataset, as shown in this embodiment of the invention. As can be seen from the figures, the embeddings generated by the graph decoupled network exhibit a clearer clustering structure in space, with nodes of the same category being more concentrated and the boundaries between different categories being clearer. Compared with the baseline models, the graph decoupled network significantly reduces overlap at category boundaries, and the node distribution exhibits stronger separability and aggregation. This visualization result qualitatively confirms the superiority of the graph decoupled network in learning discriminative node representations.

[0058] 2. Visualization of node similarity in different spaces: To visually demonstrate the ability of the graph decoupled network (Ours) of this invention to represent node category structures in the latent factor space, this invention compares it with four representative decoupled graph neural network models (DisenGCN, IPGDN, LGD-GCN, and ADGCN). To simplify the process, the first 20 nodes from the Cora test set are selected as the research object. Specifically, a node pair similarity matrix in the category space is first constructed based on the true labels: the similarity is 1 when two nodes belong to the same category, and 0 otherwise. Then, the cosine similarity between the node representations learned by each model in the latent factor space is calculated to obtain the similarity matrix in the latent factor space.

[0059] Figure 4 The visualization results of the latent factor space node similarity matrix obtained by different models on the Cora test set are presented. Figure 4 The category space similarity matrix shown in (a) can be considered an idealized reference. This matrix characterizes the purest intra-class consistency and inter-class differences. In contrast, DisenGCN, IPGDN, LGD-GCN, and ADGCN still exhibit some degree of category aliasing in the latent factor space. The latent representation learned by the graph decoupling network (f) in this embodiment of the invention is structurally closer to the category space: the similarity between intra-class nodes is significantly improved, while the similarity between inter-class nodes is significantly reduced. Therefore, the graph decoupling network in this embodiment of the invention forms a clearer category structure in the latent factor space, enabling the node representation to more accurately reflect the true category relationships.

[0060] 3. Visualization of decoupling of potential factors: Factor-level similarity modeling is performed on the latent factor spaces learned by the graph decoupling network of this embodiment and various baseline models to verify the effectiveness of latent factor decoupling. By constructing corresponding similarity matrices and performing visualization analysis, the association patterns between nodes in each latent factor space are revealed. First, the nodes in the test set are sorted according to their true category labels; then, the cosine similarity between node pairs is calculated in each latent factor space to generate the corresponding similarity matrix.

[0061] Figure 5 The visualization results show the decoupling of latent factors for different models on the Cora dataset. Figure 5 The shades of color reflect the similarity between nodes, with darker shades indicating higher similarity and lighter shades indicating lower similarity. As shown in the figure, each baseline model reveals, to some extent, the entanglement between different categories in different latent factor spaces, manifested as high similarity across categories, intertwined diagonal blocks, and unclear boundaries. In contrast, the graph decoupling network (Ours) of this embodiment, overall, shows that the seven categories of the Cora dataset can be significantly distinguished in any latent factor space, forming a clear diagonal block structure in all latent factor spaces, with significantly reduced similarity in off-diagonal regions, indicating its effective ability to distinguish different categories at the latent factor level. In other words, the high intra-class similarity and low inter-class similarity of nodes in each latent factor space indicate that it achieves good intra-class aggregation and inter-class separation at the latent factor level. This result further verifies the effectiveness of the dynamic multi-factor hierarchical decoupling module proposed in this embodiment in achieving latent factor decoupling, and also shows that the factor-level discriminative enhancement module can significantly improve category separability in each latent factor space.

[0062] 4. Visualization of latent feature correlations: To measure the correlation between different underlying factors, the component confusion matrix C is calculated: , in, The group average similarity between the i-th and j-th latent factors reflects the model's ability to effectively decouple latent factors: the lower the similarity, the weaker the correlation between the two latent factors, and the stronger the model's performance in feature decoupling.

[0063] Figure 6 This section presents a visualization of the component confusion matrix results for different models on the Cora dataset. Figure 6As can be seen, compared with other baseline models, the graph decoupling network (Ours) of this embodiment exhibits lower correlation between the representations learned in different latent factor spaces. This indicates that the graph decoupling network of this embodiment can more effectively reduce the interdependence between different latent factors, thereby achieving more complete decoupling of latent factors and maintaining higher independence of each latent factor in the latent factor space.

[0064] 5. Factor-level prototype contrastive loss embedding visualization: To verify the role of factor-level prototype contrastive loss in enhancing the class separability of each latent factor space, t-SNE visualization was performed on the node embeddings of each latent factor space. Figure 7 This paper demonstrates the distribution of nodes in each latent factor space under two scenarios: without and with factor-level prototype contrastive loss. It can be observed that without this loss, nodes of different categories overlap significantly in each latent factor space, resulting in blurred class boundaries. However, after introducing factor-level prototype contrastive loss, nodes in each latent factor space are more concentrated within their respective categories, the intervals between different categories increase significantly, and the class boundaries become clearer. This result indicates that factor-level prototype contrastive loss can effectively improve the discriminativeness of each latent factor space, promoting intra-class compactness and inter-class separability, thereby enabling the model to achieve a clearer classification structure at the latent semantic level. This further verifies the effectiveness of the factor-level discriminative enhancement module of the graph decoupling network in this embodiment of the invention in improving the separability of categories in each latent factor space.

[0065] Furthermore, the method further includes: adjusting the hyperparameters on multiple benchmark datasets. Sensitivity analysis was performed.

[0066] Hyperparameters in the overall optimization objective L on multiple benchmark datasets Sensitivity analysis was conducted to systematically evaluate the robustness and performance sensitivity of the graph decoupling network in this embodiment of the invention under different parameter configurations. Specifically, using the configuration corresponding to the best performance as a baseline, while keeping other parameters unchanged, each individual parameter was set to its baseline value. This allows us to verify the performance changes of the model across different orders of magnitude.

[0067] Figure 8 This paper presents the ablation experiment results of the graph decoupling network according to an embodiment of the present invention on six benchmark datasets. Figure 8 It can be seen that the graph decoupling network exhibits good stability under different weight parameter configurations, indicating that the model has strong robustness to hyperparameter settings. Regarding hyperparameters... The model performs well on most datasets. It remains relatively stable within the range of 0.01 to 10, while when When the dataset is too large, the accuracy of some datasets (such as Computers) drops significantly, indicating that excessively high diversity loss weights weaken the optimization effect of the main task. Regarding hyperparameters... The model exhibits relatively small performance fluctuations over a wide range, with a more stable overall trend, indicating that the factor-level prototype contrast loss has a moderate impact on model performance within a reasonable range. Overall, the model maintains high accuracy under multiple orders of magnitude variations in the two hyperparameters, verifying the robustness and generalization ability of the graph decoupling network in this embodiment of the invention regarding parameter selection.

[0068] Figure 9 A schematic diagram of the graph structure data classification device based on dynamic multi-factor decoupling according to an embodiment of the present invention is shown. Figure 9 As shown, the device includes: The acquisition module 901 is used to acquire graph data and an undirected graph constructed based on the graph data. The undirected graph includes a node set and an edge set consisting of multiple nodes. Each node is composed of multiple independent latent factors, and each latent factor is assigned a latent factor space. The hierarchical decoupling module 902 is used to input the graph data into the graph decoupling network and perform coarse-grained dynamic multi-factor decoupling and fine-grained dual constraints to perform hierarchical decoupling of multiple potential factors of the node from coarse to fine. The fine-grained dual constraints include applying structural constraints and geometric constraints to the potential factor space based on the coarse-grained decoupling representation to decouple at the fine-grained level. After applying the structural constraints, the embedding update representation of the node in the potential factor space is obtained. The geometric constraints are imposed through geometric diversity loss. The first determining module 903 is used to determine the factor-level prototype contrast loss based on the embedded update representation to enhance intra-class aggregation and inter-class separation. The second determining module 904 is used to obtain the category prediction result of the node based on a preset classifier, and to determine the cross-entropy loss based on the category prediction result and the category label of the node. Module 905 is used to construct an overall optimization objective by combining the cross-entropy loss, the geometric diversity loss, and the factor-level prototype contrast loss; The classification module 906 is used to optimize the parameters of the graph decoupled network based on the overall optimization objective, so as to classify nodes based on the optimized graph decoupled network.

[0069] The embodiments of the graph structure data classification device based on dynamic multi-factor decoupling are basically the same as the embodiments of the graph structure data classification method based on dynamic multi-factor decoupling described above, and can be referred to the above embodiments.

[0070] Figure 10The diagram shows a structural schematic of an embodiment of the computer device of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computer device.

[0071] like Figure 10 As shown, the computer device may include: a processor 402, a communications interface 404, a memory 406, and a communications bus 408.

[0072] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408. Communication interface 404 is used to communicate with other computer devices, such as clients or other server network elements. The processor 402 executes program 410, specifically performing the relevant steps described above in the computer device embodiment.

[0073] Specifically, program 410 may include program code, which includes computer-executable instructions.

[0074] Processor 402 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computer device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0075] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0076] Specifically, program 410 can be called by processor 402 to cause the computer device to perform the following operations: Acquire graph data and an undirected graph constructed based on the graph data. The undirected graph includes a set of nodes and an edge set consisting of multiple nodes. Each node is composed of multiple independent latent factors, and each latent factor is assigned a latent factor space. The graph data is input into a graph decoupling network, and coarse-grained dynamic multi-factor decoupling and fine-grained dual constraints are performed to decouple multiple potential factors of the node in a hierarchical manner from coarse to fine. The fine-grained dual constraints include applying structural constraints and geometric constraints to the potential factor space based on the coarse-grained decoupling representation to decouple at the fine-grained level. After applying the structural constraints, the embedding update representation of the node in the potential factor space is obtained. The geometric constraints are imposed through geometric diversity loss. The factor-level prototype contrast loss is determined based on the embedded update representation to enhance intra-class aggregation and inter-class separation; The class prediction result of the node is obtained based on a preset classifier, and the cross-entropy loss is determined based on the class prediction result and the class label of the node. The overall optimization objective is constructed by combining the cross-entropy loss, the geometric diversity loss, and the factor-level prototype contrast loss. The parameters of the graph decoupled network are optimized based on the overall optimization objective, so as to perform node classification based on the optimized graph decoupled network.

[0077] In one alternative approach, the fine-grained dual constraint includes applying structural and geometric constraints to the latent factor space based on the coarse-grained decoupling representation, respectively, to achieve decoupling at the fine-grained level, including: For each of the latent factor spaces, a corresponding K-nearest neighbor graph is constructed. The construction of the corresponding K-nearest neighbor graph includes: determining the Euclidean distance between the nodes based on the coarse-grained decoupled representation; determining the adjacency matrix of the nodes based on the Euclidean distance; constructing the K-nearest neighbor graph based on the coarse-grained decoupled representation and the adjacency matrix; applying a graph convolutional network to the K-nearest neighbor graph to perform an aggregation operation, obtaining the embedding update representation of the nodes in the latent factor space; and concatenating the embedding update representations of all latent factor spaces to obtain a compact decoupled representation. Based on the coarse-grained decoupling representation, the statistical response vectors of each node in different potential factor spaces are extracted. The statistical response vectors are normalized to obtain a normalized response matrix. The subspace volume spanned by the statistical response vectors is calculated based on the normalized response matrix. A geometric diversity regularization term is constructed based on the subspace volume, and the regularization term is used as the geometric diversity loss.

[0078] In one alternative approach, determining the factor-level prototype contrastive loss based on the embedded updated representation includes: Based on the embedding update representation of similar nodes in the training set, a category prototype is constructed. : , in, Let be the class prototype of class c in the m-th potential factor space. The set of nodes belonging to class c in the training set. Let u be the embedding update representation of the node in the m-th potential factor space; Calculate the normalized cosine similarity between the node and the category prototype; The prototype contrast loss is determined based on the normalized cosine similarity. : , in, The set of nodes in the training set. For the normalized cosine similarity, The category label for node u. Temperature coefficient; Prototype contrast loss across all potential factor spaces Perform a joint summation to obtain the factor-level prototype contrast loss.

[0079] In one optional approach, the coarse-grained dynamic multi-factor decoupling includes: mapping the node features to multiple potential factor spaces, calculating hub scores and authority scores based on the similarity between the node and its neighboring nodes in different potential factor spaces, determining aggregated representations of each potential factor space based on the hub scores, the authority scores, and dynamic differential weights, and concatenating the aggregated representations of all potential factor spaces to obtain the coarse-grained decoupling representation.

[0080] In one alternative approach, constructing the overall optimization objective by combining the cross-entropy loss, the geometric diversity loss, and the factor-level prototype contrast loss includes constructing the overall optimization objective L using the following formula: , in, For the cross-entropy loss, For the geometric diversity loss, This refers to the factor-level prototype contrast loss. Hyperparameters are used to balance the weights of various losses.

[0081] In an alternative approach, the method further includes: adjusting the hyperparameters on multiple benchmark datasets. Sensitivity analysis was performed.

[0082] In an alternative approach, the method further includes: Visualization operations are performed during node classification based on the optimized graph decoupled network. These visualization operations include visualization of node embeddings, visualization of node pair similarity in different latent factor spaces, visualization of latent factor decoupling, visualization of latent feature correlation, and visualization of factor-level prototype contrastive loss embeddings.

[0083] This invention provides a computer-readable storage medium storing at least one executable instruction that, when executed on a computer device, causes the computer device to perform any of the above-described method embodiments.

[0084] This invention provides a computer program that can be invoked by a processor to cause a computer device to execute any of the above-described method embodiments.

[0085] This invention provides a computer program product, which includes a computer program stored on a computer-readable storage medium. The computer program includes program instructions that, when executed on a computer, cause the computer to perform any of the above-described method embodiments.

[0086] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, the embodiments of the present invention are not directed to any particular programming language. It should be understood that the content of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0087] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0088] Similarly, it should be understood that, in order to streamline the invention and aid in understanding one or more of the various aspects of the invention, features of the embodiments of the invention are sometimes grouped together in a single embodiment, figure, or description thereof in the above description of exemplary embodiments of the invention. However, this disclosure should not be construed as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim.

[0089] Those skilled in the art will understand that modules in the computer device of the embodiments can be adaptively modified and placed in one or more computer devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or computer device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0090] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.

Claims

1. A graph structure data classification method based on dynamic multi-factor decoupling, characterized in that, The method includes: Obtain graph data and an undirected graph constructed based on the graph data. The undirected graph includes a set of nodes and an edge set consisting of multiple nodes. Each node is composed of multiple independent latent factors, and each latent factor is assigned a latent factor space. The graph data is input into a graph decoupling network, and coarse-grained dynamic multi-factor decoupling and fine-grained dual constraints are performed to decouple multiple potential factors of the node in a hierarchical manner from coarse to fine. The fine-grained dual constraints include applying structural constraints and geometric constraints to the potential factor space based on the coarse-grained decoupling representation to decouple at the fine-grained level. After applying the structural constraints, the embedding update representation of the node in the potential factor space is obtained. The geometric constraints are imposed through geometric diversity loss. The factor-level prototype contrast loss is determined based on the embedded update representation to enhance factor-level intra-class aggregation and inter-class separation; The class prediction result of the node is obtained based on a preset classifier, and the cross-entropy loss is determined based on the class prediction result and the class label of the node. The overall optimization objective is constructed by combining the cross-entropy loss, the geometric diversity loss, and the factor-level prototype contrast loss. The parameters of the graph decoupled network are optimized based on the overall optimization objective, so as to perform node classification based on the optimized graph decoupled network.

2. The method according to claim 1, characterized in that, The fine-grained dual constraint includes applying structural and geometric constraints to the latent factor space based on the coarse-grained decoupling representation, respectively, to achieve decoupling at the fine-grained level, including: For each of the latent factor spaces, a corresponding K-nearest neighbor graph is constructed. The construction of the corresponding K-nearest neighbor graph includes: determining the Euclidean distance between the nodes based on the coarse-grained decoupled representation; determining the adjacency matrix of the nodes based on the Euclidean distance; constructing the K-nearest neighbor graph based on the coarse-grained decoupled representation and the adjacency matrix; applying a graph convolutional network to the K-nearest neighbor graph to perform an aggregation operation, obtaining the embedding update representation of the nodes in the latent factor space; and concatenating the embedding update representations of all latent factor spaces to obtain a compact decoupled representation. Based on the coarse-grained decoupled representation, the statistical response vectors of each node in different potential factor spaces are extracted. The statistical response vectors are normalized to obtain a normalized response matrix. Based on the normalized response matrix, the volume of the subspace spanned by the statistical response vectors is calculated. Based on the subspace volume, a geometric diversity regularization term is constructed, and the regularization term is used as the geometric diversity loss.

3. The method according to claim 1 or 2, characterized in that, The determination of the factor-level prototype contrast loss based on the embedded update representation includes: Based on the embedding update representation of similar nodes in the training set, a category prototype is constructed. ; ， in, Let be the class prototype of class c in the m-th potential factor space. The set of nodes belonging to class c in the training set. Let u be the embedding update representation of the node in the m-th potential factor space; Calculate the normalized cosine similarity between the node and the category prototype; The prototype contrast loss is determined based on the normalized cosine similarity. ; ； in, The set of nodes in the training set. For the normalized cosine similarity, The category label for node u. Temperature coefficient; Prototype contrast loss across all potential factor spaces Perform a joint summation to obtain the factor-level prototype contrast loss.

4. The method according to claim 1, characterized in that, The coarse-grained dynamic multi-factor decoupling includes: mapping the node features to multiple potential factor spaces, calculating hub scores and authority scores based on the similarity between the node and its neighboring nodes in different potential factor spaces, determining the aggregated representation of each potential factor space based on the hub scores, the authority scores, and dynamic differential weights, and concatenating the aggregated representations of all potential factor spaces to obtain the coarse-grained decoupling representation.

5. The method according to claim 1, characterized in that, The overall optimization objective is constructed by combining the cross-entropy loss, the geometric diversity loss, and the factor-level prototype contrast loss, including constructing the overall optimization objective L using the following formula: ， in, For the cross-entropy loss, For the geometric diversity loss, This refers to the factor-level prototype contrast loss. Hyperparameters are used to balance the weights of various losses.

6. The method according to claim 5, characterized in that, The method further includes: adjusting the hyperparameters on multiple benchmark datasets. Sensitivity analysis was performed.

7. The method according to claim 1, characterized in that, The method further includes: Visualization operations are performed during node classification based on the optimized graph decoupled network. These visualization operations include visualization of node embeddings, visualization of node pair similarity in different latent factor spaces, visualization of latent factor decoupling, visualization of latent feature correlation, and visualization of factor-level prototype contrastive loss embeddings.

8. A graph structure data classification device based on dynamic multi-factor decoupling, characterized in that, The device includes: The acquisition module is used to acquire graph data and an undirected graph constructed based on the graph data. The undirected graph includes a node set and an edge set consisting of multiple nodes. Each node is composed of multiple independent latent factors, and each latent factor is assigned a latent factor space. The hierarchical decoupling module is used to input the graph data into the graph decoupling network, perform coarse-grained dynamic multi-factor decoupling and fine-grained dual constraints, so as to perform hierarchical decoupling of multiple potential factors of the node from coarse to fine. The fine-grained dual constraints include applying structural constraints and geometric constraints to the potential factor space based on the coarse-grained decoupling representation, so as to decouple at the fine-grained level. After applying the structural constraints, the embedding update representation of the node in the potential factor space is obtained. The geometric constraints are constrained by geometric diversity loss. The first determining module is used to determine the factor-level prototype contrast loss based on the embedded update representation, so as to enhance intra-class aggregation and inter-class separation; The second determining module is used to obtain the category prediction result of the node based on a preset classifier, and to determine the cross-entropy loss based on the category prediction result and the category label of the node. The construction module is used to combine the cross-entropy loss, the geometric diversity loss, and the factor-level prototype contrast loss to construct an overall optimization objective; The classification module is used to optimize the parameters of the graph decoupled network based on the overall optimization objective, so as to classify nodes based on the optimized graph decoupled network.

9. A computer device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores at least one executable instruction, which, when executed on a computer device, causes the computer device to perform the method as described in any one of claims 1-7.

Images