Fuzzy label data classification method of double-encoder graph neural network

Abstract

The invention provides a fuzzy label data classification method based on a double-encoder graph neural network, and the method comprises the steps: firstly carrying out the enhancement of input graph data, generating two different views, and inputting the two different views into a double-encoder network composed of a main encoder and an auxiliary encoder; four key node representations from different views and different encoders are obtained by forward propagation. A multi-objective composite loss function is designed, clustering consistency loss is fused in the function, and intra-class aggregation is enhanced by shortening a class prototype distance between node representation and dynamic calculation; comparing loss by unmarked nodes, and learning invariant features by maximizing the consistency of cross-view representation of the same node; and soft classification loss to process fuzzy tags by aligning model outputs with soft targets reflecting tag uncertainty. And through the loss, joint optimization and momentum updating of network parameters are carried out, and finally, robust and high-precision classification of fuzzy label graph data is realized.

Description

Technical Field

[0001] This invention belongs to the field of fuzzy label data classification technology, and particularly relates to a fuzzy label data classification method using a dual-encoder graph neural network. Background Technology

[0002] In existing technologies, graph neural networks (GNNs) have been widely used in node classification tasks, such as paper classification in citation networks and user classification in social networks. These methods typically rely on supervised learning using high-quality node labels. However, in practical applications, labels obtained through automatic annotation, crowdsourcing, and other methods often contain ambiguity and noise; for example, a node may be assigned multiple possible labels, or the labels themselves may be incorrect.

[0003] The closest existing technology is the traditional Graph Convolutional Network (GCN) and its variants. These methods directly use noisy labels for training, leading to a severe degrade in model performance because they lack mechanisms for evaluating and correcting label quality. Another approach is to use a two-stage method of "cleaning the labels first, then training the model," but this method suffers from error accumulation and the cleaning process is separated from model training, making end-to-end optimization impossible.

[0004] In summary, the main problems with existing technologies are: sensitivity to label noise and lack of a mechanism for dynamically evaluating and correcting label quality during model training, resulting in insufficient robustness and classification accuracy in fuzzy label scenarios. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of insufficient robustness and classification accuracy in fuzzy label scenarios in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A fuzzy label data classification method based on a dual-encoder graph neural network includes the following steps: S1. Obtain graph data , where V is the set of nodes, E is the set of edges, X is the node feature matrix, and Y is the initial set of node labels; S2. Based on the initial node label set Y, a candidate label matrix is ​​constructed and normalized by introducing random perturbation to generate a fuzzy label matrix that reflects the uncertainty of the labels. S3. Perform random edge dropping and / or random feature masking enhancement operations on the graph data to generate enhanced views g1 and g2; S4. Input the augmented views g1 and g2 into the dual encoder graph neural network model for forward propagation to obtain the prediction representation, target representation and classification logic value; S5. Calculate the composite loss function based on the predicted representation, the target representation, and the classification logistic value. The composite loss function includes: cluster consistency loss based on the similarity between the predicted representation and the class prototype calculated from the target representation, contrast loss based on the consistency between the predicted representation and the target representation of the same node under different augmented views, and soft classification loss between the classification logistic value and the fuzzy label matrix. S6. Using the backpropagation algorithm, update the parameters of the master encoder, predictor and classifier with the total loss of the composite loss function, and update each pair of corresponding parameters in the master encoder and the auxiliary encoder synchronously with the momentum update method. S7. Use a dual-encoder graph neural network model to perform forward propagation on the original graph data to obtain the classification logical value of each node, and obtain the classification result through the argmax operation.

[0007] Furthermore, the method for constructing the fuzzy label matrix based on the node label set Y in step S2 is as follows: (1) Initialize a candidate label matrix cdt_y with dimensions [N, C], where N is the total number of nodes and C is the total number of categories; (2) Traverse each node in the graph and use the real label as the candidate label; (3) Traverse the other categories in the graph besides the real label, and generate a random number in the range [0, 1] for each other category; if the random number is less than the preset perturbation probability, then the category is taken as a candidate label; (4) Obtain the candidate matrix based on the candidate labels, and perform row normalization on the candidate label matrix to obtain the fuzzy label matrix: .

[0008] Furthermore, the two methods for generating enhanced views in step S3 are as follows: (1) For the edge structure of the graph data, randomly discard some edges with a discard probability p1 to generate an enhanced edge index; (2) For the node feature matrix of the graph data, some feature dimensions are randomly set to zero with mask probability p2 to generate enhanced node features; (3) Based on the enhanced edge index and enhanced node features, generate two enhanced views g1 and g2.

[0009] Furthermore, the prediction representation and classification logic in step S4 are obtained using the following methods: (1) The enhanced view g1 is processed by the main encoder to obtain the hidden feature h1, and the predictor is used to perform feature mapping on the hidden feature h1 to obtain the predicted representation main_p1; (2) The enhanced view g2 is processed by the auxiliary encoder to obtain the hidden feature h2 as the target representation aux_h2; (3) Input the hidden features h1 and h2 into the classifier and project them through the fully connected layer to obtain the classification logic value of the view.

[0010] Furthermore, the specific steps for prediction and target representation are as follows: (1) The specific method for obtaining the predicted representation main_p1 is as follows: The node feature matrix g1.x and edge index g1.edge_index of the enhanced view g1 are extracted from the graph convolutional layer of the main encoder of the dual encoder graph neural network model. Based on the edge index g1.edge_index of the augmented view g1, find all direct neighbors of each node in the augmented view g1, extract the node feature vectors of the neighbors from the node feature matrix g1.x, and take the average of the node feature vectors of all neighbors of each node to obtain the aggregated neighbor feature vector of each node. The aggregated neighbor feature vector is combined with the node feature vector, and a linear transformation is performed through a learnable weight matrix to obtain the linear transformation feature. By using a nonlinear activation function, the linear transformation features are nonlinearly mapped to obtain the hidden feature h1; The hidden feature h1 is input into the predictor and mapped into the feature space to obtain the predicted representation main_p1; (2) The specific method for obtaining the predicted representation aux_h2 is as follows: The node feature matrix g2.x and edge index g2.edge_index of the enhanced view g2 are extracted from the graph convolutional layer of the auxiliary encoder of the dual encoder graph neural network model. Based on the edge index g2.edge_index of the augmented view g2, find all direct neighbors of each node in the augmented view g2, extract the node feature vectors of the neighbors from the node feature matrix g2.x, and take the average of the node feature vectors of all neighbors of each node to obtain the aggregated neighbor feature vector of each node. The aggregated neighbor feature vector is combined with the node feature vector, and a linear transformation is performed through a learnable weight matrix to obtain the linear transformation feature. By using a nonlinear activation function, the linear transformation features are nonlinearly mapped to obtain the hidden feature h2 as the target representation aux_h2.

[0011] Furthermore, the specific steps for classifying logical values ​​are as follows: (1) Input the hidden features h1 and h2 into the classifier, perform nonlinear transformation, and map them to a new feature space to obtain classification features F1 and F2; (2) Input the classification features F1 and F2 into the fully connected layer. The fully connected layer projects the feature dimensions to the number of categories C1 and C2 and obtains the classification logical values ​​logits1 and logits2.

[0012] Furthermore, the formula for calculating the cluster consistency loss in step S5 is as follows: Where: C is the total number of categories. It is the set of all nodes predicted as class c in the main encoder's predictive representation; It is the prototype center of category c, calculated as follows: ,in It is the auxiliary encoder representation of node j; It is the set of node indices predicted as class c in the auxiliary encoder representation; This represents the average cosine similarity between the two sets of vectors.

[0013] Furthermore, the formula for calculating the contrast loss of unlabeled nodes in step S5 is as follows: Wherein, main_p1 and main_p2 are the prediction representations of the two augmented views by the main encoder, and aux_h1 and aux_h2 are the target representations of the two augmented views by the auxiliary encoder.

[0014] Furthermore, the formula for calculating the soft classification loss function is as follows: ; Total loss function: in , , These represent the weights of the cluster consistency loss, the unlabeled node contrast loss, and the soft classification loss function, respectively.

[0015] Furthermore, the momentum update formula is as follows: in This represents a parameter of the main encoder; Represents the auxiliary encoder and The corresponding parameters; Momentum coefficient.

[0016] The beneficial effects of this invention are as follows: 1. Utilizing a dual-encoder structure and a fuzzy label matrix significantly improves robustness to label noise. Addressing the sensitivity of existing methods to label noise, this application constructs a fuzzy label matrix reflecting label uncertainty and employs a dual-encoder architecture consisting of a main encoder and a momentum-updated auxiliary encoder for learning. The main encoder performs rapid learning via gradient descent, while the auxiliary encoder provides a stable target representation through slow momentum updates, forming the basis for contrastive learning. This design allows the model to be trained with soft labels, avoiding overfitting to noisy labels and thus maintaining stronger stability and generalization ability even in scenarios with poor label quality.

[0017] 2. Achieving high-precision end-to-end feature learning and classification through multi-objective loss collaborative optimization. This application designs a multi-objective loss function that integrates clustering consistency loss, contrastive loss, and soft classification loss, achieving collaborative optimization within a dual-encoder framework. Clustering consistency loss utilizes the stable representation of the auxiliary encoder to calculate class prototypes, narrowing the distance between nodes of the same type; cross-view contrastive loss learns invariant features by maximizing the consistency of the representation of the same node in different views of the main and auxiliary encoders; soft classification loss aligns the model output with the fuzzy label. These three factors jointly guide the dual encoder to learn feature representations that are highly discriminative and robust to noise, overcoming the limitations of traditional methods with a single optimization objective, and achieving end-to-end performance improvement from feature learning to classification decision.

[0018] 3. Achieving superior classification performance in ambiguous label scenarios. The dual-encoder model in this application, combined with a multi-objective optimization strategy, effectively extracts discriminative features from noisy label and graph data. Experimental results show that on standard datasets, this method significantly improves classification accuracy and macro-average F1 score compared to traditional graph neural network baseline models, and the learned node representations exhibit clear intra-class clustering and inter-class separation characteristics. This verifies that the method in this application not only effectively handles label ambiguity but also fully mines data information, ultimately achieving high-precision and highly robust node classification. Attached Figure Description

[0019] Figure 1 This is a flowchart of the fuzzy label data classification method provided by the present invention; Figure 2 This is a flowchart of the triple loss calculation provided by the present invention; Figure 3 These are experimental results provided by this invention on the PubMed dataset; Figure 4 This is a t-SNE visualization diagram representing the nodes provided by the present invention; Figure 5 This is the structure diagram of the dual encoder graph neural network provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] The application principle of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0022] Please refer to Figure 1 and Figure 2 A fuzzy label data classification method based on a dual-encoder graph neural network includes the following steps: S1: Obtain graph data G=(V, E, X, Y), where V represents the set of nodes, E represents the set of edges, X represents the node feature matrix, and Y represents the initial set of node labels that may contain noise or ambiguity.

[0023] S2: Label Fuzzification: To simulate real-world scenarios or enhance model robustness, the node label set Y is perturbed with a certain probability (e.g., 0.1 to 0.2) to construct a fuzzy label matrix reflecting label uncertainty. Details are as follows:

[0024] (1) Node label set Y, where Y is a one-dimensional tensor with a length equal to the total number of nodes in the graph N, and each element y_i is the true class index of node i (for example, for a 3-class problem, y_i can take the values ​​0, 1 or 2).

[0025] (2) Initialization: Initialize a candidate label matrix cdt_y with all zeros and dimensions [N, C], where C is the total number of categories in the graph data. The candidate label matrix will be used to record all possible candidate labels for each node.

[0026] (3) Perturbation and matrix construction: Traverse each node i in the graph (from i=0 to i=N-1).

[0027] Preserving the true label: First, the true label of node i is taken as its most likely candidate label. Specifically, it is set to 1.0 at position cdt_y[i, y_i]. This ensures that the true label information of the node is preserved during the fuzzing process, and it is the most important part of its candidate set.

[0028] Randomly introduce noise labels: For node i, traverse all other categories c except its true label y_i (where c ≠ y_i).

[0029] For each other category c, generate a random number in the range [0, 1]. If the random number is less than the preset perturbation probability prob (which is typically set between 0.1 and 0.3 according to the embodiment), then category c is considered to be a possible (albeit incorrect) candidate label for node i, and the value at cdt_y[i, c] is also set to 1.0.

[0030] Through the above steps, a candidate label matrix cdt_y is generated, consisting of 0s and 1s. A value of 1 in the candidate label matrix indicates that the corresponding (node, category) is a candidate relation. For example: if... This indicates that node i belonging to category c is a candidate relation (i.e., node i could be of category c). This indicates that node i does not belong to category c.

[0031] (4) Normalization to generate fuzzy label matrix: Row normalization is performed on the candidate label matrix cdt_y to obtain the final fuzzy label matrix amb_y: S3: Graph Data Augmentation: Graph data is augmented using random edge dropping and random feature masks, generating two augmented views, g1 and g2. The specific steps are as follows:

[0032] (1) Edge enhancement operation Edges in a graph are randomly discarded to enhance their structure. This is achieved by using the edge index `edge_index` of the graph data and setting a preset discard probability. Generate an edge mask and obtain the enhanced edge index edge_index.

[0033] (2) Enhancement operations of node features Randomly mask the node features of the graph: using the node feature matrix X of the graph data, mask the nodes with a preset probability. The eigenvalues ​​of some dimensions in the node feature matrix are randomly set to zero. This operation generates a random feature mask, which is applied to the entire feature matrix to obtain the enhanced node feature matrix x.

[0034] (3) Enhanced view generation Based on the enhanced edge indices and the enhanced node feature matrix, generate enhanced views g1 and g2: S4: Forward propagation and symmetric computation are performed on the augmented views g1 and g2 using a dual-encoder graph neural network model to obtain the predicted representation, target representation, and logistic value. Please refer to [reference needed]. Figure 5 The specific steps are as follows: S401. Input augmented views g1 and g2 into the dual-encoder graph neural network model. The dual-encoder graph neural network model (BLD model) includes a master encoder, an auxiliary encoder, a predictor, a classifier, and a fully connected layer.

[0035] The predictor is a multilayer perceptron, typically consisting of two fully connected layers. It receives the output of the main encoder and maps it to another representation space.

[0036] The classifier is a multilayer perceptron, also composed of fully connected layers. It receives the output of the main encoder, maps it to the same dimension as the number of classes in the dataset, and finally outputs the probability (i.e., the classification logits) of each node belonging to each class through the Softmax function.

[0037] A fully connected layer is a simple linear layer that receives the output of the classifier and is used to adjust the final classification logic value. In some implementations, this layer may be incorporated into the classifier.

[0038] S402. Input the augmented view g1 and the augmented view g2 into the dual encoder graph neural network model for forward propagation to obtain the predicted representation main_p1 and the predicted representation aux_h2.

[0039] (1) The specific method for obtaining the predicted representation main_p1 is as follows: The node feature matrix g1.x and edge index g1.edge_index of the enhanced view g1 are extracted from the graph convolutional layer of the master encoder of the dual encoder graph neural network model.

[0040] Based on the edge index g1.edge_index of the augmented view g1, find all direct neighbors of each node in the augmented view g1. Extract the node feature vectors of the neighbors from the node feature matrix g1.x. Take the average of the node feature vectors of all neighbors of each node to obtain the aggregated neighbor feature vector of each node. This vector reflects the structural information of the local neighborhood of the node.

[0041] The aggregated neighbor feature vectors are combined with the node feature vectors, and then linearly transformed using a learnable weight matrix to obtain the linearly transformed features.

[0042] By introducing nonlinear expressive power into the model through a nonlinear activation function (such as PReLU), the linear transformation features are nonlinearly mapped to obtain a high-level, dense vector representation of all nodes in the enhanced view g1, i.e., the hidden feature h1. This representation incorporates the local topological structure information of the graph.

[0043] Predictor Mapping: The hidden feature h1 is fed into the predictor (a multilayer perceptron, MLP). This MLP typically contains fully connected layers, activation functions, and batch normalization layers. Its purpose is to map the hidden feature h1 from the feature space to another representation space used for contrastive learning. The final output of the predictor is main_p1. This is the final predicted representation from the main encoder based on the augmented view g1, which will be used for comparison with the target representation from the auxiliary encoder.

[0044] (2) The specific method for obtaining the predicted representation aux_h2 is as follows: The node feature matrix g2.x and edge index g2.edge_index of the enhanced view g2 are extracted from the graph convolutional layer of the auxiliary encoder (which has the same structure as the main encoder) of the dual encoder graph neural network model.

[0045] Based on the edge index g2.edge_index of the augmented view g2, find all direct neighbors of each node in the augmented view g2, extract the node feature vectors of the neighbors from the node feature matrix g2.x, and take the average of the node feature vectors of all neighbors of each node to obtain the aggregated neighbor feature vector of each node. This vector reflects the structural information of the local neighborhood of the node.

[0046] The aggregated neighbor feature vectors are combined with the node feature vectors, and then linearly transformed using a learnable weight matrix to obtain the linearly transformed features.

[0047] By introducing nonlinear expressive power into the model through a nonlinear activation function (such as PReLU), the linear transformation features are nonlinearly mapped to obtain a high-level, dense vector representation of all nodes in the enhanced view g2, i.e., the hidden feature h2. This representation incorporates the local topological structure information of the graph.

[0048] Direct output (no predictor): The output hidden feature h2 of the auxiliary encoder is not mapped by a predictor. The hidden feature h2 is directly output as the target representation aux_h2.

[0049] S403. The classifier based on the dual encoder graph neural network model outputs logical values ​​for classification.

[0050] Logical values ​​refer to the raw classification score vector generated by the dual-encoder network model for each node, without being normalized by the Softmax function. The dimension of this vector is equal to the number of classes in the dataset, and each element in the vector represents the raw evidence weight for that node belonging to the corresponding class.

[0051] The generation of logistic values ​​and the generation of contrastive learning representations are two parallel paths within the dual-encoder network model. Specifically:

[0052] (1) The hidden features h1 and h2 are fed into the classifier. The classifier performs a nonlinear transformation on the hidden features h1 and h2, mapping them to a new feature space that is more suitable for the final classification task, and obtains classification features F1 and F2.

[0053] (2) Input the classification features F1 and F2 into the fully connected layer. The fully connected layer projects the feature dimensions onto the number of categories C1 and C2 and outputs the result directly. The output tensor with dimensions [number of nodes, number of categories] is the classification logical value logits1 and logits2.

[0054] S5: Multi-target loss calculation: The model obtains four key outputs through forward propagation: main_p1 (the main encoder based on the view) Figure 1 (predictive representation), aux_h2 (auxiliary encoder based on view) Figure 2 (representation of), main_p2 (main encoder based on view) Figure 2 (predictive representation), aux_h1 (auxiliary encoder based on view) Figure 1 (Representation). Based on these outputs, a composite loss function consisting of three parts is calculated, as shown in the appendix. Figure 2 As shown:

[0055] S501, Cluster Consistency Loss: (1) Calculate class prototype: For each predicted class, find all nodes in the training nodes that are predicted to belong to that class, calculate the mean of the nodes on the auxiliary view representation (such as aux_h2), and use it as the prototype center of that class.

[0056] (2) Calculate similarity: Calculate the cosine similarity between the predicted representation of the main encoder (e.g., main_p1) and the corresponding class prototype center. The calculation formula is as follows: in: This represents the dot product of the master encoder's prediction representation a and the corresponding class prototype center b. Let represent the Euclidean norms (i.e., the magnitudes of the vectors) of the main encoder's predicted representation 'a' and the corresponding class prototype center 'b', respectively. 'n' is the dimension of the vector.

[0057] The loss aims to minimize the distance between a node representation and its class prototype, and can be expressed as: The mean of the values ​​is calculated, and a normalized mean is applied to all categories.

[0058] Minimizing the distance between a node representation and its class prototype is equivalent to maximizing their cosine similarity. Therefore, the cosine similarity needs to be transformed into a minimization objective. The transformation formula is as follows:

[0059] Where: C is the total number of categories. It is the set of all nodes predicted as class c in the master encoder's prediction representation. It is the prototype center of category c, calculated as follows: ,in It is the auxiliary encoder representation of node j. It is the set of node indices predicted as class c in the auxiliary encoder representation. It calculates the average cosine similarity between two sets of vectors.

[0060] S502, Unlabeled Node Contrast Loss: This is a cross-view contrastive loss. It calculates the cosine similarity between main_p1 and aux_h2, and between main_p2 and aux_h1. The loss aims to maximize the consistency of the representation of the same node across different augmented views, expressed by the formula:

[0061] .

[0062] S503, Soft classification loss: Processing fuzzy labels: Normalize the initial candidate label matrix to obtain the soft label matrix (fuzzy label matrix), i.e., amb_y = cdt_y / sum(cdt_y).

[0063] The cross-entropy loss between the classification logistic value output by the model and the soft label matrix is ​​calculated using the following formula: This loss keeps the model's learning process aligned with a dynamically updated soft objective that reflects label uncertainty.

[0064] S504, Total Losses: .

[0065] S6: Model parameter update: The parameters of the master encoder, predictor, and classifier are updated using the calculated total loss through backpropagation. This invention employs a gradient-based backpropagation algorithm to update the parameters of the master encoder, predictor, and classifier. The specific steps are as follows:

[0066] S601, Gradient Calculation (Backpropagation) (1) Initialization: Before backpropagation begins, the optimizer clears the gradient buffers of all trainable parameters to zero.

[0067] (2) Chain rule differentiation: Call the `loss.backward()` function of the deep learning framework. This function automatically starts from the total loss scalar and, according to the chain rule, calculates the partial derivative (i.e., gradient) of each trainable parameter with respect to the total loss layer by layer from the output end (total loss) to the input end (model parameters) of the computation graph. Specifically:

[0068] Calculate the gradient of the total loss with respect to each term in the composite loss function.

[0069] The gradients are passed sequentially through the soft classification loss, contrastive loss, and clustering loss functions, and then backpropagated to their inputs, namely logits, main_p1, aux_h2, main_p2, and aux_h1.

[0070] Next, the gradients continue to propagate back through the modules that generate these outputs: the gradients of logits are backpropagated through the classifier and the main encoder; the gradients of main_p1 and main_p2 are backpropagated through the predictor and the main encoder. The .grad attribute of each trainable parameter (weights and biases) in the output main encoder, predictor, and classifier will store its corresponding gradient value, indicating "the direction and strength of fine-tuning this parameter to reduce the total loss".

[0071] (3) Parameter update (optimizer step) Apply optimization rules: Call the optimizer's optimizer.step() function. The optimizer updates all trainable parameters according to the preset rules and the calculated gradients.

[0072] Update Formula: Taking the most commonly used AdamW optimizer as an example, its update process is more complex than simple stochastic gradient descent, involving momentum and adaptive learning rate, but the core idea can be simplified as follows: in: Represents any trainable parameter. It is the learning rate. This is the final update direction calculated by the optimizer based on the gradient. In this step, only the parameters of the main encoder, predictor, and classifier are updated.

[0073] The momentum update method uses the parameters of the main encoder to update the parameters of the auxiliary encoder. The momentum coefficient is usually set between 0.001 and 0.01.

[0074] The momentum update method is as follows: (1) Momentum updates are usually performed after each gradient descent update of the parameters of the master encoder.

[0075] (2) For each pair of corresponding parameters in the master encoder and the auxiliary encoder (e.g., the weights of the first layer of the master encoder and the weights of the first layer of the auxiliary encoder), perform the following operations: in: This represents a parameter of the main encoder. Represents the auxiliary encoder and The corresponding parameters. The momentum coefficient is a hyperparameter, typically set between 0.9 and 0.999.

[0076] Stability: Since m is close to 1, the parameters of the auxiliary encoder change very slowly and smoothly, which provides a stable target for contrastive learning and is the key to the successful convergence of the model.

[0077] Lag: The parameters of the auxiliary encoder are always "lagging" behind those of the main encoder. They reflect the "average state" of the parameters of the main encoder over a period of training time, rather than its latest state.

[0078] Iterative training: Repeat steps S4 to S6 until the preset maximum number of training rounds is reached.

[0079] S7: Model Inference and Output: After training, fix the parameters of the main encoder and classifier. Use them to perform forward propagation on the complete original graph data (without augmentation) to obtain the classification logistic value of each node. Finally, output the final predicted class through the argmax operation.

[0080] Example 2: The method in Example 1 will be described in detail using the application on the PubMed citation network dataset as an example.

[0081] Data Preparation and Initialization: The PubMed dataset was used. This dataset contains 19,717 scientific document nodes, 44,338 citation relationship edges, and each node has a 500-dimensional feature vector, with a total of 3 disease categories. 30% of the node labels were perturbed with a probability of 0.3 to construct a fuzzy label matrix. A two-layer GCN was initialized as the main encoder and auxiliary encoder.

[0082] Training cycle configuration: The maximum number of training epochs is set to 2000. The optimizer used is AdamW, and the learning rate is set to 5e. -4 The momentum update factor for the auxiliary encoder is set to 0.005.

[0083] Data augmentation: Obtain augmentation parameters for the PubMed dataset from the predefined agmt_dict. Based on the pre-experiment configuration, configure p1=0.5 and p2=0.2 for view g1; and p1=0.6 and p2=0.1 for view g2.

[0084] Forward Propagation and Loss Calculation: In each training epoch: Forward propagation is performed to obtain two sets of symmetrical predicted and auxiliary representations, and loss is calculated according to the method in Example 1. The three losses are added together with preset weights (1:1:1 in this example) to obtain the total loss. The total loss is backpropagated to update parameters such as the main encoder. The auxiliary encoder is updated with a momentum coefficient of 0.005.

[0085] Please refer to Figure 3 After training, the results are evaluated on the test set. Figure 3 As shown in b, the model's accuracy and macro-F1 score rise rapidly during training and eventually stabilize at a high level. Ultimately, the model achieves an accuracy of 87.1% on the PubMed dataset, significantly higher than baseline methods such as traditional GCN (approximately 82.1%), and significantly outperforms the baseline methods. Simultaneously, the training loss curve converges smoothly. Figure 3 a) On the PubMed dataset, the model's total loss value decreases steadily and tends to stabilize with increasing training epochs, indicating that the multi-objective loss function and its optimization method proposed in this invention can effectively guide model convergence. t-SNE visualization ( Figure 4 The displayed nodes indicate clear inter-class separation. Figure 4 The t-SNE visualization results show that the node representations learned by the method of this invention exhibit a clear clustering effect in two-dimensional space. Nodes of different categories are clearly divided into different regions, with compact intra-cluster nodes and separated inter-cluster nodes. (Confusion matrix attached) Figure 3 (c) This further demonstrates that the model maintains high recognition accuracy across all categories. The model's prediction accuracy across all categories in the PubMed dataset is high, with confusion primarily occurring between a few semantically similar categories. Furthermore, the values ​​on the main diagonal of the matrix (correctly classified samples) are absolutely dominant. This indicates that the method of this invention not only has high overall accuracy but also exhibits relatively balanced recognition performance across different categories, achieving a macro-average F1 score of 86.5%, further confirming its effectiveness.

[0086] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for fuzzy label data classification based on a double-encoder graph neural network, characterized in that, The method comprises the following steps: S1, obtaining graph data wherein V is a node set, E is an edge set, X is a node feature matrix, and Y is an initial node label set. S2, based on the initial node label set Y, a fuzzy label matrix reflecting label uncertainty is generated by introducing random disturbance to construct and normalize the candidate label matrix; S3, the graph data is subjected to random edge discarding and / or random feature mask enhancement operation to generate enhanced views g1 and g2; S4, the enhanced views g1 and g2 are input into the double-encoder graph neural network model for forward propagation to obtain the predicted representation, the target representation and the classification logic value; S5, based on the predicted representation, the target representation and the classification logic value, a composite loss function is calculated; the composite loss function comprises: a clustering consistency loss based on the similarity between the predicted representation and the class prototype calculated from the target representation, a contrast loss based on the consistency between the predicted representation and the target representation of the same node under different enhanced views, and a soft classification loss between the classification logic value and the fuzzy label matrix; S6, the total loss of the composite loss function is used to update the parameters of the main encoder, the predictor and the classifier through the back propagation algorithm, and each pair of corresponding parameters in the main encoder and the auxiliary encoder is synchronously updated through the momentum update method; S7, the double-encoder graph neural network model is used for forward propagation of the original graph data to obtain the classification logic value of each node, and the classification result is obtained through the argmax operation.

2. The method of claim 1, wherein the method is implemented by a dual-encoder graph neural network. The method for constructing the fuzzy label matrix based on the node label set Y in step S2 is as follows: (1) initialize a candidate label matrix cdt_y with a dimension of [N, C], wherein N is the total number of nodes and C is the total number of categories; (2) traverse each node in the graph, and take the real label as the candidate label; (3) traverse other categories in the graph except the real label, and generate a random number in the range of [0, 1] for each other category; When the random number is less than the preset disturbance probability, the category is taken as the candidate label; (4) obtain the candidate matrix based on the candidate label, and perform row normalization on the candidate label matrix to obtain the fuzzy label matrix: 。 3. The method of claim 1, wherein the method is based on a dual-encoder graph neural network. The method for generating two enhanced views in step S3 is as follows: (1) randomly discard part of the edges of the edge structure of the graph data with a discard probability p1 to generate enhanced edge indexes; (2) randomly set part of the feature dimensions to zero in the node feature matrix of the graph data with a mask probability p2 to generate enhanced node features; (3) generate two enhanced views g1 and g2 based on the enhanced edge indexes and the enhanced node features.

4. The method of claim 1, wherein the method is based on a dual-encoder graph neural network. The method for obtaining the predicted representation and the classification logic value in step S4 is as follows: (1) process the enhanced view g1 using the main encoder to obtain the hidden feature h1, and use the predictor to map the hidden feature h1 to obtain the predicted representation main_p1; (2) process the enhanced view g2 using the auxiliary encoder to obtain the hidden feature h2 as the target representation aux_h2; (3) input the hidden features h1 and h2 into the classifier and project them through the fully connected layer to obtain the classification logic value of the view.

5. The method of claim 4, wherein the method is based on a dual-encoder graph neural network. The specific steps of the predicted representation and the target representation are as follows: (1) the method for obtaining the predicted representation main_p1 is as follows: According to the graph convolution layer of the main encoder of the double-encoder graph neural network model, the node feature matrix g1.x and the edge index g1.edge_index of the enhanced view g1 are extracted; According to the edge index g1.edge_index of the enhanced view g1, all direct neighbors of each node of the enhanced view g1 are found, the node feature vectors of the neighbors are extracted from the node feature matrix g1.x, and the node feature vectors of all neighbors of each node are averaged to obtain the aggregated neighbor feature vector of each node; The aggregated neighbor feature vector is combined with the node feature vector, and linear transformation is performed through a learnable weight matrix to obtain linear transformation features; The linear transformation features are nonlinearly mapped through a nonlinear activation function to obtain hidden features h1; The hidden features h1 are input into the predictor to obtain prediction representation main_p1 through feature space mapping; (2) The method for obtaining the prediction representation aux_h2 is specifically as follows: According to the graph convolution layer of the auxiliary encoder of the double-encoder graph neural network model, the node feature matrix g2.x and the edge index g2.edge_index of the enhanced view g2 are extracted; According to the edge index g2.edge_index of the enhanced view g2, all direct neighbors of each node of the enhanced view g2 are found, the node feature vectors of the neighbors are extracted from the node feature matrix g2.x, and the node feature vectors of all neighbors of each node are averaged to obtain the aggregated neighbor feature vector of each node; The aggregated neighbor feature vector is combined with the node feature vector, and linear transformation is performed through a learnable weight matrix to obtain linear transformation features; The linear transformation features are nonlinearly mapped through a nonlinear activation function to obtain hidden features h2 as target representation aux_h2.

6. The method of claim 4, wherein the method is based on a double-encoder graph neural network. The specific steps of the classification logic value are as follows: (1) The hidden features h1 and the hidden features h2 are input into the classifier to perform nonlinear transformation and map them to a new feature space to obtain classification features F1 and F2; (2) The classification features F1 and F2 are input into a fully connected layer, which projects the feature dimension to the number of categories C1 and C2, and obtains classification logic values logits1 and logits2.

7. The dual-encoder graph neural network-based fuzzy label data classification method according to claim 1, characterized in that, The clustering consistency loss calculation formula in step S5 is as follows: where: C is the total number of classes, is the set of all nodes in the primary encoder prediction representation that are predicted to be of class c; is the prototype center of class c, computed as where is the auxiliary encoder representation of node j; is the set of node indices in the auxiliary encoder representation that are predicted to be of class c; is the average cosine similarity between the two sets of vectors.

8. The dual-encoder graph neural network-based fuzzy label data classification method according to claim 1, characterized in that, The unlabeled node pair comparison loss calculation formula in step S5 is as follows: Wherein, main_p1 and main_p2 are the prediction representations of the two enhanced views by the main encoder, and aux_h1 and aux_h2 are the target representations of the two enhanced views by the auxiliary encoder.

9. The dual-encoder graph neural network-based fuzzy label data classification method according to claim 1, characterized in that, The soft classification loss function calculation formula is as follows: ； The total loss function is as follows: wherein , , respectively represent the weight values of the clustering consistency loss, the unlabeled node contrast loss, and the soft classification loss function.

10. The method of claim 1, wherein the method is based on a dual-encoder graph neural network. The momentum update formula is as follows: wherein a certain parameter representative of the main encoder; a certain parameter representative of the main encoder; a corresponding parameter; momentum coefficient.

Images