A structure-uncertainty-oriented graph link revision method and system

Abstract

The present application relates to the technical field of graph data processing, and is especially a structure-uncertainty-oriented graph link correction method and system, which comprises taking an observed link matrix as a noisy measurement, encoding non-task structure changes through an interference factor, and encoding task effective link evidence through a task-related latent variable; based on the two, a link correction score is calculated, and is fused with the observed link matrix to generate a corrected link matrix or link set; the corrected link is used for downstream task prediction, and the model parameters are optimized jointly with the task prediction loss, the structure consistency loss and the interference regular term. The present application can simultaneously realize task link completion and non-task interference suppression when the observed graph has missing links, false links or weakly related links, output an interpretable corrected graph link result, reduce the error information transmission caused by structure uncertainty, and improve the adaptability and robustness of downstream tasks.

Description

Technical Field

[0001] This invention relates to the field of graph data processing technology, specifically to a method and system for correcting graph links in the face of structural uncertainty. Background Technology

[0002] Graph-structured data is widely found in scenarios such as social networks, information dissemination networks, knowledge graphs, reference networks, sensor networks, transportation networks, and molecular structures. In these scenarios, the links between nodes are often used by downstream models as the computational basis for information transmission, relationship inference, or classification prediction. For example, Graph Neural Networks (GNNs), by performing message passing on the observation graph and aggregating information about neighboring nodes, have become a core model for processing this type of data. However, the performance of GNNs is highly dependent on the quality of the observation graph. In practical applications, observation graphs are often incomplete and noisy: they may contain missing links, spurious links, weakly correlated links, and structural uncertainties such as sampling bias or human perturbation. Directly using such observation graphs for message passing can lead to the propagation of erroneous information and the loss of effective information, thereby reducing the accuracy of downstream tasks.

[0003] To alleviate the aforementioned problems, some researchers have attempted to introduce variational inference into graph structure learning. For example, US Patent Application Publication No. US20240386266A1 discloses a method for learning the structure of a graph neural network for medical decision-making based on task-related graph optimization. This method samples the control parameters of the graph refinement function from the variational distribution, selects the parameters that maximize the performance of downstream tasks, and thus removes some edges from the initial graph. It is effective in specific medical decision-making application scenarios.

[0004] However, this method has limitations when applied to structurally uncertain environments where missing links, spurious links, and weakly related links coexist. It only supports unidirectional edge removal, and the graph refinement function, based on a threshold of Euclidean distance between node features, can only delete edges with large distances, failing to fill in task-related links lost due to sampling bias or measurement omissions. In real-world scenarios, missing and spurious links often coexist, making a method that can only delete but not add links ineffective in restoring a complete and accurate graph structure. Furthermore, this method uses a single type of latent variable, failing to distinguish between "task-related link evidence" and "non-task-interference structures." Structural changes in the observed graph may originate from task-useful information or irrelevant noise; conflating all structural changes can easily lead to noise information being incorrectly encoded as task features, or task-useful signals being misjudged as noise and deleted. Finally, the training objective of this method is solely to maximize the performance score of downstream tasks, lacking regularization constraints on the graph structure itself. This may cause the refined graph to overfit a specific task, deviating from reliable topological evidence in the observed graph, and also failing to guarantee effective suppression of interference factors. The refined graph output by this method is only used as internal input to the GNN and does not output interpretable graph link correction results externally. The system cannot know which links have been deleted or which links should be added, and it is also difficult to use the corrected graph structure for other downstream analyses. Summary of the Invention

[0005] The purpose of this invention is to provide a graph link correction method and system for structural uncertainty, which can solve the problem of how to achieve bidirectional graph link correction that can both complete task-related links and suppress non-task interference, and output interpretable correction results when there is structural uncertainty in the observation graph where there are both missing links and spurious links.

[0006] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a graph link correction method for structural uncertainty, comprising the following steps: S1: Obtain the observation map ,in, For a set of nodes, The node feature matrix, For the set of observation links, and by Constructing the observation link matrix ; S2: Based on and input features used for interference modeling Through the interference factor encoding function Encoding yields interference factor , Used to represent non-task structure changes in the observation graph; S3: Based on , as well as Through task latent variable encoding function Encoding yields task-related latent variables , Used to represent link evidence that is useful for downstream tasks; S4: According to and Calculate the link correction score matrix ; S5: According to and Generate the corrected link matrix Or the corrected link set ; S6: Will Input downstream task model To obtain the prediction results ; S7: According to With task tags Construct a task prediction loss and combine it with structural consistency loss and interference regularization term to jointly update the interference factor encoding parameters, task latent variable encoding parameters, link correction parameters and downstream task model parameters; S8: Output the corrected graph linking results.

[0007] Preferably, the non-task structural changes include at least one of missing links, spurious links, weakly correlated links, sampling bias, measurement errors, or structural perturbations.

[0008] Preferably, the interference factor The calculation formula is as follows: ; ; in, This represents the interference factor encoding function. The parameter index of the encoding function; and Let represent the mean and log-standard deviation of the posterior distribution of the interference factor, respectively; This represents the input features used for interference modeling; This represents a random perturbation sampled from a standard normal distribution; Represents the identity matrix.

[0009] Preferably, the task-related latent variables The calculation formula is as follows: ; ; in, This represents the task latent variable encoding function. The parameter index of the encoding function; and Let represent the mean and log-standard deviation of the posterior distribution of the task-related latent variables, respectively. This represents a random perturbation sampled from a standard normal distribution.

[0010] Preferably, the link correction score matrix The calculation formula is as follows: ; in, It is the sigmoid function; and It is a positive temperature parameter; This represents element-wise matrix multiplication. This represents the pairwise interactions between task-related latent variables. This indicates evidence of task linking between node pairs; This represents the pairwise interaction between interfering factors. This indicates the degree to which the link to this node can be explained by interfering factors; Used to suppress node pairs with strong interference interpretation.

[0011] Preferably, the modified link matrix The calculation formula is: ; in, Indicates the fusion function; For preset constants, learnable parameters, node-pair gating values ​​or attention weights; The corrected link set The calculation formula is: ; in, Determine a threshold for the link; The number of nodes; and They represent the first and the One node; This indicates the corrected link status or weight.

[0012] Preferably, the prediction result The calculation formula is: ; loss function The calculation formula is: ; in: Indicates the predicted loss of the task; This represents the loss of structural consistency. Indicates the interference regularization term. and These are non-negative weight parameters; The calculation formula is as follows: ; in, It can be cross-entropy loss, mean squared error, ranking loss, or other supervision losses corresponding to downstream tasks; Tag for task; The calculation formula is as follows: ; in, For node pairs ( , ) confidence weights; Represents the binary cross-entropy; For nodes and The observation link status or observation link weight between them; The calculation formula is as follows: ; in, It is the Kullback-Leibler divergence; The parameter is The variational posterior distribution; Task-related latent variables The prior distribution; These are the weighting coefficients of the L1 norm terms; These are the weighting coefficients of the Frobenius norm term; For task latent variables With interference factors The cross-correlation matrix between them.

[0013] Preferably, the task-related latent variables A Gaussian mixture distribution is used for modeling, and differentiability is achieved through Gumbel-Softmax relaxation.

[0014] Preferably, the graph linking results include at least one of the following: adding links, deleting links, retaining links, weakening links, strengthening links, link weight, or link credibility.

[0015] Secondly, the present invention provides a graph link correction system for structural uncertainties, comprising: The graph data acquisition module is used to acquire observation graphs and construct observation link matrices. and node feature matrix ; Interference factor encoding module, used for based on and Obtain interference factor ; Task latent variable encoding module, used for... , and Obtain task-related latent variables ; Link correction module, used to... and Obtain the link correction score matrix and according to and Obtain the corrected link matrix Or the corrected link set ; Task calibration training module, used for... Perform downstream task prediction and jointly optimize task prediction loss, structural consistency loss and interference regularization term; The correction result output module is used to output the corrected graph link results; The parameter storage and update module is used to store model parameters, thresholds, temperature parameters, loss weights, and training states.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention defines graph link correction as posterior structural inference under structural uncertainty and task calibration. The observation link matrix serves only as a constraint for noisy measurements, rather than a true value that must be fully restored. Therefore, non-task structures such as spurious links and weakly correlated links in the observation graph are not incorrectly preserved, preventing noise from polluting node representations through message passing. Simultaneously, when task-related links are missing in the observation graph, these links can be actively inferred and supplemented based on node characteristics and task supervision signals, achieving task-oriented adaptive correction.

[0017] 2. This invention achieves complete bidirectional link correction in an environment of structural uncertainty by decoupling the model of task latent variables and interference factors, which can both enhance task-related links (i.e., fill in missing links) and suppress non-task interference (i.e., delete or weaken false or weakly related links).

[0018] 3. This invention integrates task link enhancement and interference link suppression into the same calculation mechanism through the link correction score matrix formula, thereby achieving bidirectional correction of missing link completion and noisy link reduction.

[0019] 4. This invention uses task prediction loss, structural consistency loss and interference regularization term for joint training, so that the correction result is both subject to task supervision and calibration, and does not deviate from the reliable topological evidence in the observation graph, thus taking into account both prediction performance and structural credibility.

[0020] 5. This invention outputs a linked matrix or linked set after adding, deleting, weakening, or strengthening, which can be directly loaded and used by downstream systems. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method steps in this invention; Figure 2 This is a schematic diagram of the overall process of the TCGR method in this invention; Figure 3 This is a schematic diagram of the link correction mechanism for decoupling task-related latent variables and interference factors in this invention; Figure 4 This is a schematic diagram of the image link correction operation in this invention; Figure 5 This is a performance comparison chart of the method GCN+TCGR in this invention and the benchmark GCN on five graph classification datasets; Figure 6 This is a schematic diagram of the graph link correction system structure in this invention; Figure 7 This is a performance comparison chart of the GCN+TCGR method in this invention, the benchmark GCN, and the robust baseline method on the Cora node classification task. Figure 8 This is a performance comparison chart of the GCN+TCGR method in this invention, the benchmark GCN, and the robust baseline method on the CiteSeer node classification task. Figure 9 This is a two-dimensional t-SNE visualization result of the Cora node representation in this invention; Figure 10 This is a confusion matrix diagram based on SeizureNet in the epilepsy classification task of this invention; Figure 11 This is a confusion matrix diagram based on self-supervised DCRNN in the epilepsy classification task of this invention; Figure 12 This is a confusion matrix diagram based on the three-tower Transformer in the epilepsy classification task of this invention; Figure 13 This is a confusion matrix diagram based on TCGR+GGN in the epilepsy classification task of this invention; Figure 14 This is a visualization of the two-dimensional t-SNE results for epilepsy representation based on SeizureNet, self-supervised DCRNN, three-tower Transformer, and TCGR+GGN in this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The English terms used in the following embodiments and their corresponding Chinese explanations, as well as their application in this invention, are summarized as follows: GCN: Graph Convolutional Network; it uses the observation adjacency matrix to normalize and aggregate the features of neighboring nodes, and is the most commonly used basic graph neural network model; as the original downstream backbone model, it observes the impact of structural perturbations on ordinary graph models.

[0024] GCN+Graph Link Correction (GCN+TCGR): A graph convolutional network after incorporating the TCGR method of this approach; the graph link correction module first obtains the corrected graph structure, and then the corrected graph structure is input into the same GCN; used to compare the downstream effects before and after correction.

[0025] GIN: Graph Isomorphic Network; it updates node representations through summation aggregation and multilayer perceptron, and has a strong ability to distinguish graph structures; it serves as a strong expressive baseline for graph classification tasks.

[0026] GraphSAGE: Neighbor Sampling Aggregation Graph Network; it samples and aggregates the neighbors of nodes, suitable for representation learning on large-scale graphs; used to test the comparative effects of different graph neural network backbones.

[0027] GAT: Graph Attention Network; it assigns attention weights to different neighbors instead of aggregating neighbors with fixed weights; it is used to compare the performance of attention aggregation models under structural perturbations.

[0028] DGCNN: Deep Graph Convolutional Network; it extracts node representations through graph convolution and uses sorted pooling to form graph-level representations; it is used as a traditional contrast model for graph classification scenarios.

[0029] DiffPool: Differentiable graph pooling network; learns soft assignment from nodes to clusters to achieve hierarchical graph representation learning; used to compare hierarchical pooling graph classification models.

[0030] RGCN: Robust Graph Convolutional Network; introduces robustness modeling during node representation or propagation to reduce the impact of abnormal and noisy edges; serves as a robust graph learning baseline under structural perturbations.

[0031] GCN-Jaccard: A graph defense method based on Jaccard similarity; suspicious edges are removed based on node feature similarity before GCN is executed; used to compare with structural defense methods based on preprocessing edge removal.

[0032] Pro-GNN: A graph structure joint learning method; it jointly learns graph structures and GNN parameters, and incorporates constraints such as sparsity and low rank; it is used to compare with existing graph structure learning / repair methods.

[0033] betaGCN: Uncertainty Graph Neural Network; enhances the stability of graph models under perturbation or noise conditions by modeling uncertainty; used to compare with robust graph neural network methods.

[0034] Example 1: TCGR (Task-Calibrated Graph Restructuring) method, also known as the task calibration graph link correction method.

[0035] Please see Figures 1-4 A graph link correction method for structural uncertainty includes the following steps: S1: Obtain the observation map The observation graph can be a directed graph, undirected graph, weighted graph, heterogeneous graph, dynamic graph, or functional connectivity graph estimated from time-series signals. Among them: For a set of nodes, , The number of nodes; The node feature matrix, , , For node feature dimensions; For the set of observed links, denoted by node The observed connections between them; and by Constructing the observation link matrix , or .

[0036] Specifically, for an undirected and unweighted graph, if the node and If there is an observation link, then set Otherwise, it is 0; for a directed graph, if the node and There is an observation link between them, set 1, otherwise 0; for a weighted graph, if the node and If there is an observation link and the link has a weight, assign the link weight to... , For nodes and The observation link status or observation link weight between them.

[0037] S2: Based on and input features used for interference modeling Through the interference factor encoding function Encoding yields interference factor , Used to represent non-task structure changes in the observation graph; Interference factor The calculation formula is as follows: ; ; ; in, This represents the interference factor encoding function. These are the parameters of the encoding function; and Let represent the mean and log-standard deviation of the posterior distribution of the interference factor, respectively; This represents the input features used for interference modeling, which can be the original node features. Constant features, random node features, or other structural auxiliary features; This represents a random perturbation sampled from a standard normal distribution; Represents the identity matrix; For the interference factor dimension.

[0038] The technical advantage of this step is that it separately represents non-task structural changes from the observation links, preventing subsequent link correction processes from mistaking noisy links, sampling biases, or weakly correlated edges as valid task links. Non-task structural changes include at least one of the following: missing links, spurious links, weakly correlated links, sampling biases, measurement errors, or structural perturbations.

[0039] As another implementation method, interference factor It can be calculated by graph neural networks, graph convolutional networks, attention networks, multilayer perceptrons, variational encoders, or probabilistic graphical models.

[0040] S3: Based on , as well as Through task latent variable encoding function Encoding yields task-related latent variables , Used to represent link evidence that is useful for downstream tasks; Task-related latent variables The calculation formula is as follows: ; ; ; in, This represents the task latent variable encoding function. The parameter index of the encoding function; and Let represent the mean and log-standard deviation of the posterior distribution of the task-related latent variables, respectively. This represents a random perturbation sampled from a standard normal distribution; For task latent variable dimensions.

[0041] Task labels during training phase Instead of being used as direct input to the encoder during testing, Z is calibrated for its learning direction through subsequent task losses. This step ensures that link corrections are no longer determined solely by observed topology or node similarity, but rather form potential link evidence that can be calibrated by downstream tasks.

[0042] As another implementation method, task-related latent variables It can be implemented using single Gaussian, Gaussian mixture, low-rank decomposition, attention representation, contrastive learning representation, or other differentiable latent variable forms.

[0043] S4: According to and Calculate the link correction score matrix ; Link to correct score matrix The calculation formula is as follows: ; ; in, This is the sigmoid function, used to map link scores to the interval [0, 1]. and These are positive temperature parameters, used to adjust the smoothness of task-linked evidence and interference explanations, respectively. This represents element-wise matrix multiplication. This represents the pairwise interactions between task-related latent variables. This indicates evidence of task linking between node pairs; This represents the pairwise interaction between interfering factors. This indicates the degree to which the link to this node can be explained by interfering factors; Used to suppress node pairs with strong interference interpretation. Represents a node and The link validity score determined by both task latent variables and interference factors. The larger the value, the more likely it is to be a node. and The more links are formed, preserved, or strengthened between them.

[0044] The above formula does not rely on restoring the observation link matrix. Instead of targeting the link correction score, it defines it as the coupling result of task link evidence and interference suppression term. Therefore, when an observed link is mainly explained by interference factors, even 1. Its correction score can also be reduced; when an unobserved link has strong task evidence, even if If the value is 0, its correction score can also be increased.

[0045] As another implementation method, linking to correct scores The calculation can be performed using inner product, bilinear function, multilayer perceptron, attention scoring function or kernel function, as long as it simultaneously utilizes task-related latent variables and interference factors for link enhancement and interference suppression.

[0046] S5: According to and Generate the corrected link matrix Or the corrected link set ; Corrected link matrix The calculation formula is: ; in, Or a binary matrix obtained by thresholding; The fusion function can be any one of fixed-weight fusion, learnable-weight fusion, node-pair-gated fusion, or attention-gated fusion. This can be a preset constant, a learnable parameter, a node pair's relevant gating value, or an attention weight; when When the value is large, the correction results depend more on the observed graph; when When the value is smaller, the correction result depends more on the link correction score obtained from task calibration.

[0047] If the application scenario requires a discrete set of links, then a threshold can be used. Will Convert to a corrected set of links, corrected set of links The calculation formula is: ; in, Determine a threshold for the link; The number of nodes; and They represent the first and the One node; This indicates the corrected link status or weight.

[0048] Under this transformation, if 0 and The corresponding operation is to complete the link; if 1 and The corresponding operation is to delete or weaken the link; if 1 and If so, the corresponding operation is to preserve or enhance the link; if 0 and If so, then no link will be maintained.

[0049] In directed graph scenarios, and They can be calculated separately; in undirected graph scenarios, they can be calculated separately. Symmetry processing is performed. In the weighted graph scenario, It can be directly used as the corrected link weight.

[0050] S6: Will Input downstream task model To obtain the prediction results ; Prediction results The calculation formula is: ; in, For downstream task models; The node feature matrix; This is the corrected link matrix.

[0051] S7: According to With task tags Construct a task prediction loss and combine it with structural consistency loss and interference regularization term to jointly update the interference factor encoding parameters, task latent variable encoding parameters, link correction parameters and downstream task model parameters; loss function The calculation formula is: ; in: Indicates the predicted loss of the task; This represents the structural consistency loss, used to keep the correction results compatible with reliable evidence in the observation plot; Indicates the interference regularization term. and These are non-negative weight parameters.

[0052] The calculation formula is as follows: ; in, It can be cross-entropy loss, mean squared error, ranking loss, or other supervision losses corresponding to downstream tasks; For task tags.

[0053] The calculation formula is as follows: ; in, For node pairs ( , ) confidence weights; Represents the binary cross-entropy; For nodes and The observation link status or observation link weight between them; The calculation formula is as follows: ; in, It is the Kullback-Leibler divergence; The parameter is The variational posterior distribution; Task-related latent variables The prior distribution; These are the weighting coefficients of the L1 norm terms; These are the weighting coefficients of the Frobenius norm term; For task latent variables With interference factors The cross-correlation matrix between them.

[0054] The posterior of the latent variables in the constraint task is close to the prior. The L1 norm term constrains the interpretation of disturbances to avoid excessive density, while the Frobenius norm term is used to reduce the correlation between latent variables and disturbance factors. This is achieved by minimizing... This can make the gradient of task loss... Backward propagation to link-corrected scores Task-related latent variables and interference factors This enables graph link correction calibrated by downstream tasks.

[0055] S8: Output the corrected graph linking results, which include at least one of the following: added links, deleted links, retained links, weakened links, enhanced links, link weights, or link credibility.

[0056] To verify the technical effectiveness of the method in this embodiment, the following experiments were conducted: As the reconnection rate of the graph structure increases, the classification accuracy and AUC-ROC of the baseline GCN typically decrease significantly. This phenomenon indicates that performing message passing directly on the perturbed observation graph harms model performance, thus confirming the necessity of graph link correction in environments of structural uncertainty. Please refer to Table 1 for a comparison of the baseline graph neural network and the network with the added link correction mechanism of this invention.

[0057] Table 1. Comparison results between the basic graph neural network and the network with the added link correction mechanism of this invention. The above results demonstrate that the technical effect of this invention does not come from simple superimposed graph neural networks and denoising processing, but from the link correction mechanism of task calibration: by separating task-related link evidence from interference structures and making task loss directly affect the link correction results, it is possible to reduce the transmission of erroneous information when the observation graph is disturbed or incomplete.

[0058] Please see Figure 5 The figure shows a performance comparison between the proposed GCN+TCGR method and the benchmark GCN on five graph classification datasets. Experiments were conducted on the MUTAG, PROTEINS, ENZYMES, NCI1, and AIDS graph classification datasets with perturbation rates of 0%, 10%, 30%, and 50%, respectively. As the perturbation rate increases, the performance of the benchmark GCN significantly decreases. The proposed method outperforms the benchmark GCN on all datasets and with all perturbation rates, demonstrating stronger robustness and stability. Specifically, with a 50% perturbation on the MUTAG dataset, the accuracy of the proposed method improves to 88.27%; with a 50% perturbation on the AIDS dataset, the proposed method reaches 81.19%, a significant improvement over the benchmark GCN's 75.13%.

[0059] Example 2: Correction System.

[0060] Please see Figure 6 A graph link correction system for structural uncertainties includes: The graph data acquisition module is used to acquire observation graphs and construct observation link matrices. and node feature matrix ; Interference factor encoding module, used for based on and Obtain interference factor ; Task latent variable encoding module, used for... , and Obtain task-related latent variables ; Link correction module, used to... and Obtain the link correction score matrix and according to and Obtain the corrected link matrix Or the corrected link set ; Task calibration training module, used for... Perform downstream task prediction and jointly optimize task prediction loss, structural consistency loss and interference regularization term; The correction result output module is used to output the corrected graph link results; The parameter storage and update module is used to store model parameters, thresholds, temperature parameters, loss weights, and training states.

[0061] Each module can be executed by a processor in a server, edge device, or cloud computing node.

[0062] Example 3: Modeling with a Gaussian mixture distribution.

[0063] The difference from Example 1 is that in step S3, the task-related latent variables... A Gaussian mixture distribution can be used to model and represent the posterior of the multimodal structure. The calculation formula is as follows: ; in, The parameter is The variational posterior distribution; For nodes The task-related latent variable vector; The amount of mixed components; ; Represents a node Belongs to the The probability of each potential structural component; and Let represent the mean and standard deviation of the corresponding components, respectively. When discrete component sampling is required, Gumbel-Softmax relaxation can be used to maintain training differentiability; its calculation formula is as follows: ; in, For nodes In the The soft-assignment weights of each component; For positive temperature parameters, It is Gumbel noise; then it is sampled to obtain: ; in, For nodes In the A standard normal random noise vector with 1 component; It is an identity matrix.

[0064] Example 4: Application of news dissemination image link correction scenario.

[0065] Nodes can represent news reports, content fragments, publishing accounts, dissemination events, or cross-platform content instances, while observed links can represent forwarding, citations, reposting, association with content from the same source, association based on publication time, or association based on semantic similarity. Due to factors such as missing cross-platform data collection, undocumented reposting relationships, similar content without genuine dissemination relationships, and collaborative dissemination by abnormal accounts, news dissemination graphs may also contain missing links, false links, and weakly related links. When using the method of this invention, the above dissemination relationships can be constructed into an observed graph, and a corrected dissemination link result can be generated by decoupling task-related latent variables and interference factors. This result can be used for dissemination path analysis, key node identification, abnormal dissemination structure identification, or content evolution analysis. This application is only used to illustrate a type of graph data that this invention can process and does not constitute a limitation on the scope of protection of this invention.

[0066] Example 5: Node classification application.

[0067] This embodiment uses a semi-supervised node classification task as an example to illustrate the specific application of the present invention.

[0068] Semi-supervised node classification using Cora or CiteSeer is employed. Both datasets contain sparse bag-of-words node attributes and share a citation graph as their graph structure. This setup is used to examine whether TCGR can still calibrate the message passing process after structural perturbations alter the local neighborhood, even when only some node labels are available. After training according to the steps described in Example 1, class prediction is performed on unlabeled nodes.

[0069] Please see Figure 7 For a performance comparison chart of the GCN+TCGR method, the baseline GCN, and robust baseline methods on the Cora node classification task, please refer to [link / reference]. Figure 8The graph shows a performance comparison of the GCN+TCGR method, the baseline GCN, and robust baseline methods on the CiteSeer node classification task. The robust baseline methods include Pro-GNN, GAT, RGCN, GCN-Jaccard, and betaGCN. This experiment examines whether the posterior graph inferred by TCGR can still support reliable node-level message passing when the observed reference links are contaminated. On Cora, GCN+TCGR maintains strong competitiveness across the entire perturbation range and is more effective at high reconnection ratios, where the base GCN is more sensitive to neighborhood contamination. This phenomenon is consistent with the design goal of TCGR, which is to infer missing task-related links and suppress spurious connections by refining the graph structure before downstream aggregation.

[0070] The results on CiteSeer are more complex. GCN+TCGR shows only a small advantage over robust baseline methods, and GCN-Jaccard is competitive under weak perturbation settings. This result is reasonable because CiteSeer has sparse and high-dimensional node features, and feature similarity-based edge filtering can remove a significant portion of noisy links. Therefore, we do not interpret this result as TCGR being superior in all settings, but rather as indicating that TCGR is more valuable when task supervision can provide information beyond simple feature similarity. When a certain type of perturbation can be handled well by lightweight preprocessing rules, an additional posterior graph generator may only bring small incremental benefits.

[0071] Please see Figure 9 As shown, this is a visualization of the two-dimensional t-SNE results represented by Cora nodes. Experiments show that when the observation links are randomly perturbed by 70%, the accuracy of the base GCN using the method of this invention is better than that of the GCN without the method of this invention, indicating that task calibration link correction helps to maintain a clearer class structure under high perturbation conditions.

[0072] Example 6: Application of spatiotemporal diagram classification.

[0073] This embodiment uses epileptic electroencephalogram (EEG) signal classification as an example to illustrate the application of the present invention in spatiotemporal graph classification. EEG data comes from 20 channels and is converted into a dynamic functional connectivity graph through a sliding time window. This task differs from static benchmark datasets because the graphs in these datasets are not fixed true topologies, but rather estimated connectivity patterns that change over time and with epilepsy type. In this embodiment, the estimated graph is treated as a noisy observation, and link correction is performed using the method described in Embodiment 1.

[0074] Specifically, this embodiment selects the following baseline models for comparison: SeizureNet (a CNN-based multispectral deep feature learning model), Self-Supervised DCRNN (a graph neural network improved using self-supervised learning), Three-Tower Transformer (a model based on the Three-Tower Transformer architecture), and TCGR+GGN. The TCGR+GGN method in this embodiment follows the GGN architecture, retaining its temporal encoder and spatial message passing components, and inserting TCGR as a graph remodeling module. Given temporal features extracted from EEG fragments, TCGR infers a refined graph A* and uses it for the downstream GNN. Task latent components are used to capture connection patterns that contribute to the prediction of seizure type, while perturbation factors absorb non-task-related variations, such as subject-specific effects, measurement noise, and weak information channel correlations. The model is trained with the same prediction objective as the downstream classifier, incorporating the aforementioned consistency and perturbation regularization terms.

[0075] Please see Figures 10-13 The graphs shown are the confusion matrices for the epilepsy classification task of this invention, based on SeizureNet, self-supervised DCRNN, three-tower Transformer, and TCGR+GGN, respectively. Darker colors indicate more concentrated predictions for the corresponding class, and off-diagonal areas represent class confusion. The value multiplied by 100% represents the corresponding accuracy. Experiments show that the graph structure corrected using the TCGR+GGN method in this embodiment achieves better classification accuracy than the baseline method that directly uses the estimated graph, and the confusion matrix shows reduced inter-class confusion.

[0076] Please see Figure 14 The image shows the visualization results of two-dimensional t-SNE for epilepsy representation based on SeizureNet, self-supervised DCRNN, three-tower Transformer, and TCGR+GGN. It compares the representation distributions obtained by different models, with different colors representing different categories. GGN with TCGR, supported by a task-calibrated latent graph, exhibits a more significant category separation effect. This demonstrates that graph link correction helps improve the separability of feature representations.

[0077] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A graph link correction method for structural uncertainty, characterized in that, Includes the following steps: S1: Obtain the observation map ,in, For a set of nodes, The node feature matrix, For the set of observation links, and by Constructing the observation link matrix ; S2: Based on and input features used for interference modeling Through the interference factor encoding function Encoding yields interference factor , Used to represent non-task structure changes in the observation graph; S3: Based on , as well as Through task latent variable encoding function Encoding yields task-related latent variables , Used to represent link evidence that is useful for downstream tasks; S4: According to and Calculate the link correction score matrix ; S5: According to and Generate the corrected link matrix Or the corrected link set ; S6: Will Input downstream task model To obtain the prediction results ; S7: According to With task tags Construct a task prediction loss and combine it with structural consistency loss and interference regularization term to jointly update the interference factor encoding parameters, task latent variable encoding parameters, link correction parameters and downstream task model parameters; S8: Output the corrected graph linking results.

2. The graph link correction method for structural uncertainty according to claim 1, characterized in that: The non-task structural changes include at least one of missing links, spurious links, weakly correlated links, sampling bias, measurement errors, or structural perturbations.

3. The graph link correction method for structural uncertainty according to claim 1, characterized in that: The interference factor The calculation formula is as follows: ； ； in, This represents the interference factor encoding function. The parameter index of the encoding function; and Let represent the mean and log-standard deviation of the posterior distribution of the interference factor, respectively; This represents the input features used for interference modeling; This represents a random perturbation sampled from a standard normal distribution; Represents the identity matrix.

4. The graph link correction method for structural uncertainty according to claim 1, characterized in that: The task-related latent variables The calculation formula is as follows: ； ； in, This represents the task latent variable encoding function. The parameter index of the encoding function; and Let represent the mean and log-standard deviation of the posterior distribution of the task-related latent variables, respectively. This represents a random perturbation sampled from a standard normal distribution.

5. The graph link correction method for structural uncertainty according to claim 1, characterized in that: The link correction score matrix The calculation formula is as follows: ； in, It is the sigmoid function; and It is a positive temperature parameter; This represents element-wise matrix multiplication. This represents the pairwise interactions between task-related latent variables. This indicates evidence of task linking between node pairs; This represents the pairwise interaction between interfering factors. This indicates the degree to which the link to this node can be explained by interfering factors; Used to suppress node pairs with strong interference interpretation.

6. The graph link correction method for structural uncertainty according to claim 1, characterized in that: The corrected link matrix The calculation formula is: ； in, Indicates the fusion function; For preset constants, learnable parameters, node-pair gating values ​​or attention weights; The corrected link set The calculation formula is: ； in, Determine a threshold for the link; The number of nodes; and They represent the first and the One node; This indicates the corrected link status or weight.

7. The graph link correction method for structural uncertainty according to claim 1, characterized in that: The prediction results The calculation formula is: ； loss function The calculation formula is: ； in: Indicates the predicted loss of the task; This represents the loss of structural consistency. Indicates the interference regularization term. and These are non-negative weight parameters; The calculation formula is as follows: ； in, It can be cross-entropy loss, mean squared error, ranking loss, or other supervision losses corresponding to downstream tasks; Tag for task; The calculation formula is as follows: ； in, For node pairs ( , ) confidence weights; Represents the binary cross-entropy; For nodes and The observation link status or observation link weight between them; The calculation formula is as follows: ； in, It is the Kullback-Leibler divergence; The parameter is The variational posterior distribution; Task-related latent variables The prior distribution; These are the weighting coefficients of the L1 norm terms; These are the weighting coefficients of the Frobenius norm term; For task latent variables With interference factors The cross-correlation matrix between them.

8. The graph link correction method for structural uncertainty according to claim 1, characterized in that: The task-related latent variables A Gaussian mixture distribution is used for modeling, and differentiability is achieved through Gumbel-Softmax relaxation.

9. A graph link correction method for structural uncertainty according to claim 1, characterized in that: The graph linking results include at least one of the following: adding links, deleting links, retaining links, weakening links, strengthening links, link weight, or link credibility.

10. A graph link correction system for structural uncertainty, characterized in that, include: The graph data acquisition module is used to acquire observation graphs and construct observation link matrices. and node feature matrix ; Interference factor encoding module, used for based on and Obtain interference factor ; Task latent variable encoding module, used for... , and Obtain task-related latent variables ; Link correction module, used to... and Obtain the link correction score matrix and according to and Obtain the corrected link matrix Or the corrected link set ; Task calibration training module, used for... Perform downstream task prediction and jointly optimize task prediction loss, structural consistency loss and interference regularization term; The correction result output module is used to output the corrected graph link results; The parameter storage and update module is used to store model parameters, thresholds, temperature parameters, loss weights, and training states.

Images