Graph neural network node classification method based on mutual exclusion group sparse learning

Abstract

The invention relates to a graph neural network node classification method based on mutual exclusion group sparse learning, and the method comprises the following steps: introducing a mutual exclusion group sparse regularization mechanism into an existing graph neural network, and finely regulating and controlling a feature selection process through a strategy of intra-group feature sharing and inter-group feature competition; the regular term is embedded into the graph neural network, and according to the feature importance and the noise level of different network layers, the weight of feature retention and elimination and the proportion of balance group sparse and mutual exclusion sparse regularization are adaptively adjusted, so that the robustness and the stability of feature selection are enhanced; and fine control and selection of multi-scale image features are realized. In practice, by using the method, the flexibility of feature selection and the network generalization of the neural network can be improved, and meanwhile, the robustness and the expression ability of the model are remarkably enhanced.

Description

Technical Field

[0001] The present invention relates to the field of neural network node classification, and in particular to a graph neural network node classification method based on mutually exclusive group sparse learning. Background Art

[0002] Graph Neural Networks (GNNs) effectively model dependencies in graph-structured data by introducing a message-passing mechanism between nodes. They are widely used in tasks such as node classification, link prediction, and clustering. GNNs have also achieved remarkable results in fields such as biochemistry, recommender systems, traffic prediction, and computer vision. However, real-world graph data often contains a large amount of noise and redundant features, which severely hamper the learning and generalization capabilities of GNNs. Although some existing methods incorporate feature selection mechanisms, their adaptability remains insufficient in complex noisy environments. Therefore, improving the robustness and feature selection capabilities of GNNs in node classification tasks has become a key research direction.

[0003] Graph Convolutional Network (GCN), as a neural network model based on graph convolution, pioneered the introduction of the concept of convolution into graph neural networks. It uses Fourier transform to convert the feature matrix of the graph from the time domain to the frequency domain, and then defines the convolution operation in the frequency domain. In this way, the application of trainable convolution operations on graph structured data is realized, which can accurately capture the local structure of the graph and the interaction between nodes. For example, in the graph structure of a social network, GCN can effectively mine the social relationship patterns between user nodes. However, the GCN algorithm based on frequency domain convolution usually requires eigendecomposition of the Laplace matrix, which makes the computational and time complexity high, and limits the efficiency when processing large-scale graph data.

[0004] The Graph Attention Network (GAT) takes a different approach and introduces an attention mechanism to replace traditional graph convolution operations. During each iteration, GAT assigns a unique attention weight to each neighboring node, and then performs a weighted summation of the node features in the neighborhood to update the feature representation of the target node. This method is like the model automatically focusing on important information among a lot of information. Taking the knowledge graph as an example, GAT can more keenly capture the degree of correlation between different entity nodes and effectively capture the complex local relationships in graph-structured data. Moreover, GAT is independent of the global structure of the graph and does not require high-cost computing operations such as matrix decomposition, thereby greatly reducing resource consumption.

[0005] Currently, a variety of effective adversarial attack methods have been designed for GNNs (GCN and GAT are collectively referred to as the GNNs series) models, but their effectiveness is limited when directly applied to various heterogeneous graph structures or complex downstream tasks. The existing AsGNN method introduces a feature selection mechanism to guide the model to automatically select the most representative attribute features during the learning process. However, its use of parameter weight normalization can weaken the differences between features, resulting in the loss of important information and affecting model performance. Another FSGNN method, by introducing a soft selector, adaptively assigns importance weights to node features of different hops to achieve feature selection. However, for features within the same hop, this method assigns uniform weights, making it difficult to capture finer-grained dynamic relationships between nodes, affecting the effectiveness of the attack. Existing feature selection methods are insufficient in capturing complex relationships between features and adapting to dynamic changes in graph data, which can easily lead to the loss of key information and reduced feature selection accuracy. Summary of the Invention

[0006] In response to the above-mentioned problems existing in the prior art, the technical problem to be solved by the present invention is: how to improve the robustness and accuracy of graph neural networks based on the mutually exclusive group sparse learning mechanism.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] A graph neural network node classification method based on mutually exclusive group sparse learning includes the following steps:

[0009] S100: Select any public standard document dataset, take a technical field in the standard document dataset as a category, perform data preprocessing on each document in the standard document dataset to obtain a document sample graph set; take the vertex of each document sample graph in the document sample graph set as a node, and select N nodes from each category as a training set;

[0010] S200: Construct the graph neural network model EGS-RGNNs. EGS-RGNNs is an improved robust graph neural network RGNNs, which adds the mutually exclusive group sparsity regularization term EGS to RGNNs.

[0011] S300: Constructing the loss function L of EGS-RGNNs egs-rgnns , the calculation expression is as follows:

[0012]

[0013] Among them, L gnn Denotes the loss function of GNNs, Ω(W (k) ) represents the mutually exclusive group sparse regularization term of the k-th layer weight matrix, k = 1, 2, ..., K represents the number of model layers;

[0014] S400: Set the minimum threshold ε and the maximum number of iterations threshold of the loss function change, use the training set as input data, and use L egs-rgnns Training EGS-RGNNs:

[0015] In each round of iteration, the weight parameter matrix of the current neural network layer is first updated by the gradient descent method, and then the group sparsity regularizer proximal operator Pgd21(W,λ,μ) and the mutually exclusive sparsity regularizer proximal operator Pgd12(W,λ,μ) are applied to regularize the weight parameter matrix in sequence; where W represents the weight matrix, λ represents the regularization parameter used to control the weight matrix W and λ>0; μ represents the balance parameter and 0≤μ≤1, balancing the ratio of group sparsity and mutually exclusive sparsity regularization;

[0016] After completing a round of iterative training, the updated weight parameter matrix is ​​substituted into the loss function to update the current loss function value L current And calculate L current and the previous round loss function value L prev The difference ΔL, that is, ΔL=|L current -L prev |; Perform several rounds of iterative training. When ΔL is less than ε or the number of training rounds reaches the threshold, stop training and obtain the trained graph neural network model EGS-RGNNs';

[0017] S500: Select a standard document dataset Y to be classified, perform data preprocessing on Y according to S100 to obtain Y', and then input Y' into EGS-RGNNs' to obtain the classification results of the documents in Y.

[0018] Preferably, in S100, the data preprocessing method is to convert all node information in the standard document dataset through the graph neural network data loader Dataloader and output the data format used by the graph neural network.

[0019] Preferably, the calculation expression of the mutually exclusive group sparse regularization term EGS in S200 is as follows:

[0020] Ω(W)=λ[μ||W|| 2,1 +(1-μ)||W|| 1,2 ]

[0021] in, Ω(W) represents the mutually exclusive group sparse regularization term of the weight matrix W.

[0022] As an advantage, in said S300, the weight matrix W (K) The calculation formula is:

[0023]

[0024] in, represents the weight matrix of the kth layer at the tth iteration, represents the intermediate state after the t-th gradient descent update, η represents the learning rate, Represents the classification loss function L gnn For weight W (k) gradient.

[0025] Preferably, in S300, when EGS-RGNNs uses GCN / GAT as the base model, the loss function is calculated as follows:

[0026]

[0027] Among them, L gcn (A, X; W) represents the loss function of AsGCN when using GCN as the base model;

[0028]

[0029] in, represents the loss function of AsGAT when GAT is used as the base model, β represents the regularization parameter used to control the attention weight vector α, represents the set of attention weights,

[0030] Preferably, in S400, the specific content of applying the group sparse regularizer proximal operator Pgd21 (W, λ, μ) and the mutually exclusive sparse regularizer proximal operator Pgd12 (W, λ, μ) to perform regularization processing on the weight parameter matrix is ​​as follows:

[0031] Next, the regularizer proximal operator is used to process the non-smooth regularization term;

[0032] For the weight matrix of the kth layer, minimize the following objective:

[0033]

[0034] in,

[0035] Ω(W)=λ[μ||W|| 2,1 +(1-μ)||W|| 1,2 ]

[0036]

[0037]

[0038] Pgd21(W,λ,μ) and Pgd12(W,λ,μ) were used to treat l 2,1 and l 1,2 norm, and then w ij Make updates;

[0039] Definition 2,1 The norm is to calculate the l2 norm of each row of the weight matrix first, and then sum the l2 norms of all rows; define l 1,2 The norm is to first calculate the l1 norm of each column of the weight matrix, and then square the calculation results of all l1 norms and add them together;

[0040] Processing 2,1 The proximal operator of the group sparsity regularizer of the norm is expressed as:

[0041]

[0042] The optimal solution of Pgd21(W,λ,μ) is simplified to:

[0043]

[0044] Among them, ||w i ||2 represents the l2 norm of the i-th feature group, is a truncation function used to adjust w i The weight of

[0045] Processing 1,2 The norm enables the proximal operator of the mutually exclusive sparse regularizer to be expressed as:

[0046]

[0047] The optimal solution of Pgd12(W,λ,μ) is simplified to:

[0048]

[0049] Among them, ||w i ||1 represents the l1 norm of the i-th feature group, |w ij | indicates w i The absolute value of the jth element of λ||w i ||1 means w i Overall punishment, Represents a truncation function for w ij Perform sparse processing;

[0050] When Pgd21(W,λ,μ) and Pgd12(W,λ,μ) are regularized, w is realized. ij Update, thereby reversely updating Ω(W) and updating the model loss function L egs-rgnns.

[0051] Compared with the prior art, the present invention has at least the following advantages:

[0052] 1. The mutually exclusive group sparsity (EGS) method proposed in this invention integrates the two regularization mechanisms of mutually exclusive sparsity and group sparsity, and adaptively adjusts the weights according to the network layer structure through the parameter μ. Group sparsity regularization promotes the retention or removal of feature groups as a whole, while mutually exclusive sparsity regularization screens key features and compresses redundancy within the group, achieving "intra-group competition and inter-group mutual exclusion". This effectively solves the problems of traditional feature selection methods being susceptible to interference from noise and redundant features, as well as extreme feature processing, accurately screens key feature groups and features, significantly improves the model's ability to model exclusive features for different tasks or categories, and enhances the model's discriminability and robustness in high-dimensional multi-task learning scenarios.

[0053] 2. Currently, most feature selection methods based on graph neural networks lack accuracy and robustness when faced with high-dimensional data and noisy environments. This paper combines the EGS method with a graph convolutional network (GCN) and a graph attention network (GAT) to construct the EGS-RGCN and EGS-RGAT models. By introducing the EGS regularization term into the loss functions of GCN and GAT, the model can automatically highlight important features and suppress the influence of noisy features during training, significantly improving the accuracy of node classification tasks, especially in high-dimensional and noisy environments, demonstrating better feature extraction capabilities and overall performance.

[0054] 3. An optimization strategy based on the proximal gradient descent (PGD) algorithm first performs conventional gradient updates on the smooth portion of the loss function, then uses a proximal operator to handle the regularization portion. Group sparsity and mutually exclusive sparsity regularizers are applied to the weight parameter matrix, and the LASSO regularizer is applied to the attention weight vector of the GAT network. This strategy addresses the issues of low training efficiency, poor prediction accuracy, and weak adaptability when processing high-dimensional data and data with noisy or redundant features. This enables the model to maintain high training efficiency and prediction accuracy even when processing complex data, improves adaptability through its robustness, and ensures stable performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a simplified flow chart of the method of the present invention;

[0056] Figure 2 To compare the node classification results of EGS-RGCN and EGS-RGAT with other baseline methods on the noisy dataset;

[0057] Figure 3 Comparison of visual classification results between different models on the noisy dataset Cora;

[0058] Figure 4 Comparison of visual classification results between different models on the noisy dataset Citeseer. DETAILED DESCRIPTION

[0059] The present invention is described in further detail below.

[0060] See also Figure 1-Figure 4 , a graph neural network node classification method based on mutually exclusive group sparse learning, comprising the following steps:

[0061] S100: Select any public standard document dataset, take a technical field in the standard document dataset as a category, perform data preprocessing on each document in the standard document dataset to obtain a document sample graph set; take the vertex of each document sample graph in the document sample graph set as a node, and select N nodes from each category as a training set;

[0062] In S100, the data preprocessing method is to convert all node information in the standard literature dataset through the graph neural network data loader Dataloader and output the data format used by the graph neural network. The graph neural network data loader Dataloader is a data loader provided by PyTorch, which is a prior art.

[0063] Processing the standard literature data involves extracting all node feature matrices, adjacency tables, and classification labels from the standard literature dataset. Using the graph neural network data loader, the raw data is converted into a model-usable format through the following core steps: 1) adding self-loop connections to enhance node relationships, 2) symmetric normalization of the adjacency matrix to balance the influence of different nodes, and 3) converting the classification labels into numerical codes. The loader automatically completes this entire process, ultimately outputting a standard data structure for graph neural network training.

[0064] S200: Construct the graph neural network model EGS-RGNNs. EGS-RGNNs is an improved robust graph neural network RGNNs, which adds the mutually exclusive group sparsity regularization term EGS to RGNNs. RGNNs are robust graph neural networks, including robust graph convolutional neural networks and robust graph attention neural networks. EGS is a regularization mechanism that integrates the mutually exclusive sparsity regularization mechanism and the group sparsity regularization mechanism. Graph convolutional neural networks, graph attention neural networks, mutually exclusive sparsity regularization mechanisms, and group sparsity regularization mechanisms are all existing technologies.

[0065] EGS, which integrates two types of sparse regularization, can encourage feature competition and achieve the purpose of intra-group feature sharing. RGCN stands for Robust GCN, where robust means robustness. This is the traditional GNNs that adds this mutually exclusive group sparse regularization during model optimization, which enhances the noise resistance of the entire model and achieves refined control of feature selection. EGS can make adaptive selections based on the structure of different network layers. The optimization formula has a parameter μ, which is the weight of the sparse regularization method. Generally speaking, in the low layers of the network, features require overall structural information, so the μ of the first layer is larger, allowing group sparse regularization to dominate. The second layer makes the features more complex and more discriminative, so reducing the mul of the second layer makes the mutually exclusive sparse regularization dominant. The precise selection of features will be reflected here.

[0066] The calculation expression of the mutually exclusive group sparse regularization term EGS in S200 is as follows:

[0067] Ω(W)=λ[μ||W|| 2,1 +(1-μ)||W|| 1,2 ]

[0068] in, Ω(W) represents the mutually exclusive group sparse regularization term of the weight matrix W.

[0069] S300: Constructing the loss function L of EGS-RGNNs egs-rgnns , the calculation expression is as follows:

[0070]

[0071] Among them, L gnn Denotes the loss function of GNNs, Ω(W (k) ) represents the mutually exclusive group sparse regularization term of the k-th layer weight matrix, k = 1, 2, ..., K represents the number of model layers;

[0072] In S300, the weight matrix W (K) The calculation formula is:

[0073]

[0074] in, represents the weight matrix of the kth layer at the tth iteration, represents the intermediate state after the t-th gradient descent update, η represents the learning rate, Represents the classification loss function L gnn For weight W (k) gradient;

[0075] In S300, when EGS-RGNNs uses GCN / GAT as the base model, the loss function is calculated as follows:

[0076]

[0077] Among them, L gcn (A, X; W) represents the loss function of AsGCN when using GCN as the base model; Feature selection guided graph convolutional neural network AsGCN is the existing technology;

[0078]

[0079] in, represents the loss function of AsGAT when GAT is used as the base model, β represents the regularization parameter used to control the attention weight vector α, represents the set of attention weights, Feature selection guided graph attention network AsGAT is an existing technology;

[0080] S400: Set the minimum threshold ε and the maximum number of iterations threshold of the loss function change, use the training set as input data, and use L egs-rgnns Training EGS-RGNNs:

[0081] In each iteration, the weight parameter matrix of the current neural network layer is first updated by the gradient descent method, and then the group sparsity regularizer proximal operator Pgd21(W,λ,μ) and the mutually exclusive sparsity regularizer proximal operator Pgd12(W,λ,μ) are applied to regularize the weight parameter matrix in sequence; wherein W represents the weight matrix, λ represents the regularization parameter used to control the weight matrix W and λ>0; μ represents the balancing parameter and 0≤μ≤1, balancing the ratio of group sparsity and mutually exclusive sparsity regularization; the group sparsity regularizer proximal operator and the mutually exclusive sparsity regularizer proximal operator are existing technologies;

[0082] After completing a round of iterative training, the updated weight parameter matrix is ​​substituted into the loss function to update the current loss function value L current And calculate L current and the previous round loss function value L prev The difference ΔL, that is, ΔL=|L current -L prev |; Perform several rounds of iterative training. When ΔL is less than ε or the number of training rounds reaches the threshold, stop training and obtain the trained graph neural network model EGS-RGNNs'; the loss function value of the previous round of the first round L prev The value calculated when the weight parameter matrix is ​​initialized;

[0083] In the S400, the specific contents of applying the group sparse regularizer proximal operator Pgd21(W,λ,μ) and the mutually exclusive sparse regularizer proximal operator Pgd12(W,λ,μ) to perform regularization processing on the weight parameter matrix are as follows:

[0084] Next, the regularizer proximal operator is used to process the non-smooth regularization term;

[0085] For the weight matrix of the kth layer, minimize the following objective:

[0086]

[0087] in,

[0088] Ω(W)=λ[μ||W|| 2,1 +(1-μ)||W|| 1,2 ]

[0089]

[0090]

[0091] Pgd21(W,λ,μ) and Pgd12(W,λ,μ) were used to treat l 2,1 and l 1,2 norm, and then w ij Make updates;

[0092] Definition 2,1 l 2,1 The norm is to calculate the l2 norm of each row of the weight matrix, also known as the Euclidean norm, and then sum the l2 norms of all rows; define l 1,2 The norm is to first calculate the l1 norm of each column of the weight matrix, that is, the sum of the absolute values ​​of the elements in the column, and then square the calculation results of all l1 norms and add them together;

[0093] Processing 2,1 The proximal operator of the group sparsity regularizer of the norm is expressed as:

[0094]

[0095] The optimal solution of Pgd21(W,λ,μ) is simplified to:

[0096]

[0097] Among them, ||w i ||2 represents the l2 norm of the i-th feature group, is a truncation function used to adjust w i The weight of

[0098] Processing1,2 The norm enables the proximal operator of the mutually exclusive sparse regularizer to be expressed as:

[0099]

[0100] The optimal solution of Pgd12(W,λ,μ) is simplified to:

[0101]

[0102] Among them, ||w i ||1 represents the l1 norm of the i-th feature group, |w ij | indicates w i The absolute value of the jth element of λ||w i ||1 means w i Overall punishment, Represents a truncation function for w ij Perform sparse processing;

[0103] When Pgd21(W,λ,μ) and Pgd12(W,λ,μ) are regularized, w is realized. ij Update, thereby reversely updating Ω(W) and updating the model loss function L egs-rgnns .

[0104] The optimal solution is simplified by using “vector soft thresholding”. ij Update the solution using the simplified form of the optimal solution, that is, the entire row w of the weight matrix i Perform uniform scaling, W, λ, μ are the parameters used for uniform scaling, and the scaling parameters are the same for each row. The truncation function is when || w i When ||2 is greater than λμ, it is not truncated. When ||w i When ||2 is less than or equal to λμ, the i-th feature group is set to 0 and will be "truncated"; λμ is the standard form of the proximal operator in the proximal gradient method, which has been theoretically proven to be the existing technology for vector soft thresholding.

[0105] For the GAT network that includes the attention mechanism, the attention weight vector α needs to be iteratively updated. The formula for its update at the tth iteration is:

[0106]

[0107] in, The attention weight vector of the kth layer at the tth iteration, represents the intermediate state after the gradient descent update at the tth iteration, η' represents the learning rate, Represents the weight vector The gradient of , dealing with the attention weight vector in the non-smooth regularization term, needs to minimize the following objectives:

[0108]

[0109] The LASSO regularizer proximal operator is used to process the l1 norm separately. The LASSO regularizer proximal operator that processes the l1 norm can be expressed as:

[0110]

[0111] Specifically, the optimal solution of Pgd1(α,β) can be simplified as:

[0112]

[0113] Among them, α represents the attention weight vector, β represents the regularization parameter used to control the sparsity of α, and |α| represents the length of the attention weight vector. Indicates the scaling ratio, controls the compression degree of α, represents a non-negative truncation function, ensuring that when When , the value is directly reduced to zero to achieve sparseness.

[0114] S500: Select a standard document dataset Y to be classified, perform data preprocessing on Y according to S100 to obtain Y', and then input Y' into EGS-RGNNs' to obtain the classification results of the documents in Y.

[0115] Experimental content and results

[0116] Dataset selection: The present invention selected Glorot, Pubmed, Cora and Citeseer datasets. For each dataset, 20 nodes were randomly selected from each class as the training set, 500 nodes as the validation set, and 1000 nodes as the test set. The EGS-RGAT model was trained with a maximum training round of 5000. If the validation loss did not decrease for 100 consecutive epochs, the training was stopped. Glorot was used to initialize all network parameters. For the Cora and Citeseer datasets, the first layer consists of 8 attention heads, each of which calculates 8 features, which are spliced ​​to generate a 64-dimensional feature representation for each node. The second layer is used for node classification, using a single attention head to output several features of the category, and the final category probability is obtained through softmax. For the Pubmed dataset, due to the small number of training samples, the model uses 8 attention heads in the second layer to adapt to the characteristics of the data.

[0117] Benchmark model: The method of the present invention selects GCN, GAT, Masked GCN, TO-GCN, BGNN, RGCN, FSGNN, rGNN, GIB, AsGCN, AsGAT and EGS-RGCN as benchmark models for comparison.

[0118] Evaluation criteria and parameter settings: The experiment selected accuracy as the core evaluation metric to measure the classification performance of the model on different datasets. Specifically, accuracy is defined as the ratio of the number of samples correctly classified by the model to the total number of samples, which can be expressed as:

[0119]

[0120] TP represents the number of positive samples correctly identified by the model, that is, the number of true positive samples successfully classified as positive by the model. TN represents the number of negative samples correctly identified by the model, that is, the number of true negative samples correctly classified as negative. FP represents the number of negative samples incorrectly identified as positive by the model, that is, the number of samples that actually belong to the negative class but are misclassified as positive by the model. FN represents the number of positive samples incorrectly identified as negative by the model, that is, the number of samples that actually belong to the positive class but are misclassified as negative by the model. Dropout is set to 0.6, and each layer of the model has separate regularization parameters λ and β, as well as a balancing parameter μ, to assist in training. These three parameters are adjusted to filter features in the weight matrix.

[0121] Multiple sets of regularization parameters λ and β, as well as the balancing parameter μ, were set for EGS-RGAT, and the best-performing set of λ, β, and μ was selected. In the EGS-RGCN model, the best classification accuracy on the Cora dataset occurred with λ1 set to 1.5–2.0 and λ2 set to 0.01–0.03. On the Citeseer dataset, the optimal range was λ1 set to 1.0–1.5 and λ2 set to 0.025–0.075. On the Pubmed dataset, the best performance was achieved with λ1 set to 1.0–2.0 and λ2 set to 0.1–1.0. Experiments showed that the best performance on the Cora, Citeseer, and Pubmed datasets occurred with μ1 set to 0.8–1.0 and μ2 set to 0.0–0.2.

[0122] In EGS-RGAT training, the optimal regularization parameter ranges are: λ1 of 0.4–0.5 and λ2 of 0.5–0.75 for the Cora dataset; λ1 of 0.4–0.6 and λ2 of approximately 2.0 for the Citeseer dataset; and λ1 of 0.5–1.0 and λ2 of approximately 0.6 for the Pubmed dataset. Regarding the balancing parameters, μ1 of 0.8–1.0 and μ2 of approximately 0.3–0.4 for the Cora and Citeseer datasets, and μ1 of 0.6–0.8 and μ2 of approximately 0.2 for the Pubmed dataset. Comparative results are shown in Tables 1 and 2.

[0123] Table 1 shows the comparison of classification results of EGS-RGCN and other benchmark methods on standard datasets

[0124]

[0125]

[0126] The method of the present invention was empirically evaluated on the node classification task on three standard citation network datasets: Cora, Citeseer, and PubMed. The experimental results show that the proposed EGS-RGCN model achieved excellent classification performance on multiple datasets. Compared with the traditional GCN model, the accuracy on the Cora dataset was improved by about two percentage points, and the overall performance was better than existing representative methods including Masked GCN and GAT. On the Citeseer dataset, the accuracy of this method exceeded that of various advanced comparison models such as TO-GCN, showing stronger robustness and adaptability. Even though the overall improvement of the model in the PubMed dataset was limited, EGS-RGCN still achieved a stable classification effect with a result variance as low as 0.1, further verifying its robustness.

[0127] The EGS-RGCN network model trained by this method can adaptively suppress redundant feature interference and accurately screen key node features when faced with high-noise graph data by introducing a mutually exclusive group sparsity regularization mechanism, effectively improving the accuracy and robustness of graph neural networks in tasks such as node classification. This method is particularly suitable for graph data mining tasks in complex and high-noise scenarios, such as social network analysis and recommender systems, and has broad application prospects and practical value.

[0128] Table 2 shows the comparison of classification results of EGS-RGAT and other benchmark models on standard datasets

[0129] Method Name Cora Citeseer Pubmed GCN 81.5 70.3 79.00 Masked GCN 82.70 72.00 79.30 TO-GCN 83.10 72.70 79.50 CAT 83.0±0.70 72.50±0.70 79.00±0.30 BGCN-T 82.0±0.10 71.90±0.00 79.40±0.10 RGCN 82.8±0.60 71.20±0.50 79.10±0.30 GIB 82.2±0.52 70.50±0.74 78.94±0.60 FSGNN 82.3±0.39 70.88±0.46 79.10±0.29 rGNN 82.6±0.25 72.22±0.90 79.20±0.29 AsGAT 84.31±0.4 74.32±0.67 79.15±0.70 EGS-RGCN 83.4±0.25 71.84±0.30 79.9±0.10 EGS-RGAT 84.78±0.36 74.75±0.56 79.52±0.13

[0130] The proposed method significantly outperforms existing mainstream graph neural network models, including GCN, GAT, and AsGAT, on three standard graph datasets: Cora, Citeseer, and PubMed. The accuracy on the Cora dataset reached 84.78%, demonstrating superior classification performance and stability. The accuracy on Citeseer was 74.75%, a significant improvement over AsGAT. Compared to the baseline model EGS-RGCN, the accuracy improved by 1.38% on Cora and 1.91% on Citeseer. It also achieved leading accuracy on PubMed, demonstrating its adaptability to large-scale graph structures.

[0131] Through the collaborative design of mutually exclusive group sparsity and attention mechanism, the proposed method can effectively filter key information and suppress redundancy and noise interference, which not only improves the classification performance, but also enhances the robustness and generalization ability of the model in complex environments.

[0132] Table 3 shows the comparison of node classification results of EGS-RGNN and EGS-RGAT with other baseline methods on the noisy dataset.

[0133]

[0134] Table 3 shows the accuracy of different graph neural network models in node classification tasks on graph datasets with added noise interference, comprehensively reflecting the robustness of each model under uncertainty. In terms of average performance, the EGS-RGCN and EGS-RGAT models outperform existing mainstream methods on all datasets. For example, on the Cora dataset, EGS-RGAT achieved an average accuracy of 83.88%, significantly exceeding AsGAT's 82.96%. On Citeseer, EGS-RGAT achieved an average accuracy of 73.63%, also outperforming the current best comparison model. On the PubMed dataset, EGS-RGCN led all methods with an average accuracy of 79.33%.

[0135] The EGS series models demonstrated consistent minimum accuracy across repeated experiments, significantly outperforming traditional models, further demonstrating their robustness and reliability. This characteristic makes the EGS-RGNNs series particularly suitable for real-world tasks involving noisy and complex graph data, such as social network analysis, recommendation systems, and node classification in biomedical graphs, demonstrating promising application prospects and adaptability.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A graph neural network node classification method based on mutually exclusive group sparse learning, characterized by: The steps include: S100: Select any public standard document dataset, take a technical field in the standard document dataset as a category, perform data preprocessing on each document in the standard document dataset to obtain a document sample graph set; take the vertex of each document sample graph in the document sample graph set as a node, and select N nodes from each category as a training set; S200: Construct the graph neural network model EGS-RGNNs. EGS-RGNNs is an improved robust graph neural network RGNNs, which adds the mutually exclusive group sparsity regularization term EGS to RGNNs. S300: Constructing the loss function L of EGS-RGNNs egs-rgnns , the calculation expression is as follows: Among them, L gnn Denotes the loss function of GNNs, Ω(W (k) ) represents the mutually exclusive group sparse regularization term of the k-th layer weight matrix, k = 1, 2, ..., K represents the number of model layers; S400: Set the minimum threshold ε and the maximum number of iterations threshold of the loss function change, use the training set as input data, and use L egs-rgnns Training EGS-RGNNs: In each round of iteration, the weight parameter matrix of the current neural network layer is first updated by the gradient descent method, and then the group sparsity regularizer proximal operator Pgd21(W,λ,μ) and the mutually exclusive sparsity regularizer proximal operator Pgd12(W,λ,μ) are applied to regularize the weight parameter matrix in sequence; where W represents the weight matrix, λ represents the regularization parameter used to control the weight matrix W and λ>0; μ represents the balance parameter and 0≤μ≤1, balancing the ratio of group sparsity and mutually exclusive sparsity regularization; After completing a round of iterative training, the updated weight parameter matrix is ​​substituted into the loss function to update the current loss function value L current And calculate L current and the previous round loss function value L prev The difference ΔL, that is, ΔL=|L current -L prev |; Perform several rounds of iterative training. When ΔL is less than ε or the number of training rounds reaches the threshold, stop training and obtain the trained graph neural network model EGS-RGNNs'; S500: Select a standard document dataset Y to be classified, perform data preprocessing on Y according to S100 to obtain Y', and then input Y' into EGS-RGNNs' to obtain the classification results of the documents in Y.

2. The graph neural network node classification method based on mutually exclusive group sparse learning according to claim 1, characterized in that: In S100, the data preprocessing method is to convert all node information in the standard document dataset through the graph neural network data loader Dataloader and output the data format used by the graph neural network.

3. The graph neural network node classification method based on mutually exclusive group sparse learning according to claim 2, characterized in that: The calculation expression of the mutually exclusive group sparse regularization term EGS in S200 is as follows: Ω(W)=λ[μ||W|| 2,1 +(1-μ)||W|| 1,2 ] in, represents the mutually exclusive group sparse regularization term of the weight matrix W.

4. The graph neural network node classification method based on mutually exclusive group sparse learning according to claim 3, characterized in that: In S300, the weight matrix W (K) The calculation formula is: Among them, W t (k) represents the weight matrix of the kth layer at the tth iteration, represents the intermediate state after the t-th gradient descent update, η represents the learning rate, Represents the classification loss function L gnn For weight W (k) gradient.

5. The graph neural network node classification method based on mutually exclusive group sparse learning according to claim 4, characterized in that: In S300, when EGS-RGNNs uses GCN / GAT as the base model, the loss function is calculated as follows: Among them, L gcn (A, X; W) represents the loss function of AsGCN when using GCN as the base model; in, represents the loss function of AsGAT when GAT is used as the base model, β represents the regularization parameter used to control the attention weight vector α, represents the set of attention weights, 6. The graph neural network node classification method based on mutually exclusive group sparse learning according to claim 5, characterized in that: In the S400, the specific contents of applying the group sparse regularizer proximal operator Pgd21(W,λ,μ) and the mutually exclusive sparse regularizer proximal operator Pgd12(W,λ,μ) to perform regularization processing on the weight parameter matrix are as follows: The regularizer proximal operator is used to process the non-smooth regularization term; For the weight matrix of the kth layer, minimize the following objective: in, Ω(W)=λ[μ||W|| 2,1 +(1-μ)||W|| 1,2 ] Pgd21(W,λ,μ) and Pgd12(W,λ,μ) were used to treat l 2,1 and l 1,2 norm, and then w ij Make updates; Definition 2,1 The norm is to calculate the l2 norm of each row of the weight matrix first, and then sum the l2 norms of all rows; define l 1,2 The norm is to calculate the l1 norm of each column of the weight matrix first, and then square the calculation results of all l1 norms and add them together; Processing 2,1 The proximal operator of the group sparsity regularizer of the norm is expressed as: The optimal solution of Pgd21(W,λ,μ) is simplified to: Among them, ||w i ||2 represents the l2 norm of the i-th feature group, is a truncation function used to adjust w i The weight of Processing 1,2 The norm enables the proximal operator of the mutually exclusive sparse regularizer to be expressed as: The optimal solution of Pgd12(W,λ,μ) is simplified to: Among them, ||w i ||1 represents the l1 norm of the i-th feature group, |w ij | indicates w i The absolute value of the jth element of λ||w i ||1 means w i Overall punishment, Represents a truncation function for w ij Perform sparse processing; When Pgd21(W,λ,μ) and Pgd12(W,λ,μ) are regularized, w is realized. ij Update, thereby reversely updating Ω(W) and updating the model loss function L egs-rgnns .

Images