Graph anomaly detection method based on low-rank contrastive learning and reconstruction
By using low-rank contrastive learning and reconstruction methods to generate low-rank views and combining them with graph neural networks, the accuracy and robustness issues of graph anomaly detection in existing technologies are solved, and effective identification of abnormal nodes and noise filtering are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2026-04-09
- Publication Date
- 2026-06-05
AI Technical Summary
Existing graph anomaly detection methods have limitations when processing graph structure data, such as the inability to effectively distinguish abnormal nodes and preserve the semantic structure of the graph. In particular, graph contrastive learning methods are susceptible to noise interference and rely on high-quality positive and negative pairs, leading to inaccurate detection.
A low-rank contrastive learning and reconstruction method is adopted. A low-rank view is generated by performing low-rank approximation on node attributes and topology. Then, a graph neural network is combined for contrastive learning and reconstruction to generate a low-rank view to filter out anomalies and noise and preserve the global information of the graph.
It effectively filters out abnormal nodes and noise, improving the accuracy and robustness of graph anomaly detection and ensuring the correctness and superiority of the detection.
Smart Images
Figure CN122156934A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of graph anomaly detection, and in particular to a graph anomaly detection method based on low-rank contrastive learning and reconstruction. Background Technology
[0002] Graph anomaly detection (GAD) aims to identify patterns that significantly deviate from normal patterns in graph-structured data. Node-level detection is the most common approach because it is suitable for entity-based anomaly assessment. Due to the unstructured and complex nature of graphs, node anomalies can be categorized into topological anomalies and attribute anomalies: topological anomalies refer to errors in node connections, while attribute anomalies indicate damage to the intrinsic features of nodes. Early unsupervised GAD methods, such as residual analysis and CUR decomposition, provided metric-based anomaly quantification but often neglected the interdependencies between nodes. The emergence of graph neural networks has facilitated the development of self-supervised methods, where graph autoencoders utilize reconstruction errors to detect anomalies.
[0003] Contrastive learning is employed to improve discriminative power by measuring the consistency between nodes and their contextual neighborhoods. Extensions such as multi-scale contrastive learning further enrich representation learning using graph diffusion or edge modification. While graph contrastive learning-based GAD methods are effective, they also have inherent limitations: i) Inherent anomalies in the graph can cause some GAD methods to inadvertently select anomalous nodes, resulting in unreliable pairs that propagate anomalies during GCN encoding. This undermines the effectiveness of GAD, as it heavily relies on high-quality positive-negative pairs. ii) Random augmentations used for view generation introduce external noise and often fail to preserve the semantic structure of the graph, further degrading GAD performance. Low-rank approximation, a core technique in signal processing for dimensionality reduction and noise suppression, has been applied to GNNs to handle graph-structured data. However, the potential of GAD has not been fully explored. To mitigate these drawbacks, we shift from local and random contrastive learning to global and semantically aware learning, proposing a low-rank contrastive learning and reconstruction method for GAD. In this method, low-rank approximations are performed on node attributes and topology to generate a global low-rank view. This effectively filters out anomalies and noise while incorporating global information into the graph contrastive learning process. Furthermore, the multi-view comparison learning and reconstruction module is designed based on low-rank prior knowledge. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings and deficiencies of existing graph contrastive learning anomaly detection schemes, and to propose a graph anomaly detection method based on low-rank contrastive learning and reconstruction. By performing low-rank approximation on node attributes and topology to generate a low-rank view, the method effectively filters out abnormal nodes and noise interference while preserving the original graph structure.
[0005] To achieve the above objectives, the technical solution provided by this invention is: a graph anomaly detection method based on low-rank contrastive learning and reconstruction, comprising the following steps:
[0006] 1) Low-rank view generation and subgraph sampling: Obtain graph data and generate low-rank graph data by dimensionality reduction through SVD singular value decomposition. Then, use the restarted random walk algorithm to sample the subgraph to obtain the original view and the low-rank view.
[0007] 2) Low-rank contrastive learning: Contrast pairs are constructed on the original view and the low-rank view to perform contrastive learning and obtain contrastive learning loss;
[0008] 3) Low-rank attribute reconstruction: Perform low-rank attribute reconstruction on the original view and the low-rank view to obtain the final reconstruction loss of the original view and the low-rank view;
[0009] 4) Model training: Train the graph neural network model by combining contrastive learning loss and reconstruction loss;
[0010] 5) Anomaly score discrimination: Based on the trained graph neural network model, anomaly discrimination is performed on the nodes in the graph under test to identify potential anomaly nodes.
[0011] Furthermore, in step 1), the detailed process of low-rank view generation and subgraph sampling is as follows:
[0012] 1.1) Obtaining Graph Data: The graph is represented in the following form: , including nodes ,in It is a node attribute matrix. It is the set of real numbers. It is the number of nodes. It is a data dimension. Representing the 1 node It is an adjacency matrix, in the adjacency matrix Chinese Represents a node and nodes Connection status, Represents a node and nodes There is a connecting edge between them. This indicates a node. and nodes There are no connecting edges between them, where ;
[0013] 1.2) SVD Decomposition: First, during the pre-training phase of the graph, random SVD decomposition is applied to the graph. and ,in and It corresponds to a matrix and The left singular matrix, and These are diagonal matrices, with their main diagonal elements representing and The singular values, and and It corresponds to a matrix and The right singular matrix is obtained; then, the first r singular values are retained to obtain the reconstructed node attributes. and adjacency matrix ,Right now and ,in and These represent the number of singular values, and yes and The low-rank approximation of , where and SVD decomposition is used to generate low-rank graph data, which is then used for the next step of subgraph sampling.
[0014] 1.3) Subgraph Sampling: The Restarted Random Walk (RWR) algorithm is used to extract subgraphs from both the original graph data and the low-rank graph data. For each node... Sample a size of The subgraphs are used to obtain the original views. and low-rank view ,in and It is the adjacency matrix and feature matrix formed by sampling subgraphs from the original graph data, and similarly... and It is the adjacency matrix and feature matrix formed by the sampled subgraphs of the low-rank graph data.
[0015] Furthermore, in step 2), the detailed process of the low-rank contrastive learning is as follows:
[0016] 2.1) Constructing the contrast loss between views First, the target node Corresponding original view and low-rank view The inputs are fed into a single GCN layer to obtain the embeddings: and ,in It is a learnable weight matrix. and Represents the target node in two views In the characteristic matrix and The feature attributes are masked, i.e., their values are set to zero; then, the final representation of the two views is obtained using the average readout function within the GCN: and ,in Indicates the node number in the two views. and express The node with the sequence number is embedded and output by GCN. and These represent the subgraph embeddings of the original view and the low-rank view, respectively. This refers to the number of nodes in the corresponding view; then, the nodes are mapped to the same space where the views are embedded using MLP, i.e., the node mapping between the two views. and ,in and Representing the target node respectively In the properties of the original view and the low-rank view, It is an MLP mapping; then, the sigmoid function is used in the discriminator to obtain the target node. Similarity scores between the subgraph embeddings corresponding to the two views and the subgraph embeddings: and ,in It is the sigmoid function. These are the weight parameters corresponding to the discriminator; following the concept of contrastive learning, the similarity scores between the two views... and To request a more precise match, it can be expressed as: ,in Represented as the Frobenius paradigm;
[0017] 2.2) Constructing the in-view contrast loss First, construct the comparison pairs needed for contrastive learning, i.e., the target nodes. The sampled subgraphs and their sampling subgraphs constitute positive sample pairs, while the target node... With another node in the same batch The sampled subgraphs form negative sample pairs; then, the similarity scores of the positive and negative sample pairs are calculated separately in the low-rank view, i.e.: and ,in, For nodes Subgraph embedding of the low-rank view corresponding to the sampled subgraph Close to 1 Approaching 0, we then use BCE loss to train the contrast loss for the low-rank view, i.e., Similarly, a similarity score can also be obtained from the original view. and The contrast loss of the original view is trained using BCE loss, i.e. Next, the final node-subgraph contrast loss, i.e., the in-view contrast loss, is calculated: ,in These are parameters that control the weights of the original view and the lower-rank view;
[0018] 2.3) Integration yields contrastive learning loss By integrating the inter-view contrast loss and the intra-view contrast loss, and adding them together, we can obtain the contrastive learning loss. .
[0019] Furthermore, in step 3), the low-rank attribute reconstruction includes the following steps:
[0020] 3.1) Calculate the reconstruction loss of the original view and the reconstruction loss of the low-rank view, i.e., the node... Reconstruction loss of the original view and nodes Reconstruction loss of the low-rank view ,in and They are and The series combination of adjacent nodes embedded in the middle It is an MLP mapping;
[0021] 3.2) Calculate the final reconstruction loss for the original view and the low-rank view: .
[0022] Furthermore, in step 4), the model training includes:
[0023] First, by combining the reconstruction loss and the contrastive learning loss, the final loss function is obtained. ,in, These are the parameters that balance the reconstruction loss and the contrastive learning loss; finally, the loss function is used. Training a graph neural network model.
[0024] Furthermore, in step 5), the abnormal score determination includes the following steps:
[0025] 5.1) Satisfying the form The graph to be tested is input into the trained graph neural network model, and steps 1) and 2) are repeated to obtain the corresponding representation, and then step 5.2) is executed.
[0026] 5.2) Calculate the anomaly score for low-rank view contrast learning. That is, the difference in similarity between negative sample pairs and positive sample pairs, expressed as: ,in The higher the score, the more abnormal the node is, and then the same abnormality score is obtained from the original view comparison learning. Finally, the outlier scores for the contrastive learning component were obtained. ;
[0027] 5.3) Calculate the reconstruction anomaly score for the low-rank view. ,in The higher the score, the more abnormal the node is, and the reconstruction anomaly score of the original view is obtained in the same way. Finally, the final anomaly score of the low-rank attribute reconstruction is obtained. ;
[0028] 5.4) Integrate the two abnormal scores above to obtain the final abnormal score of the node. The higher the score, the more abnormal the node, thus detecting potentially abnormal nodes.
[0029] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0030] 1. This invention introduces low-rank approximation into graph anomaly detection for the first time, and proposes a graph anomaly detection method while ensuring the correctness of graph anomaly detection.
[0031] 2. This invention generates low-rank views by reducing the rank of feature attributes and topology, and designs a low-rank attribute reconstruction loss as an aid, thus ensuring the robustness of graph anomaly detection.
[0032] 3. This invention realizes low-rank-based contrastive learning and reconstruction, verifies the effectiveness of the low-rank contrastive learning and reconstruction strategy, and ensures the superiority of the method of this invention for graph anomaly detection. Attached Figure Description
[0033] Figure 1 This is a schematic diagram of the logic flow of the present invention.
[0034] Figure 2 This is a framework diagram of the present invention; in the diagram, This indicates the average readout pattern of GCN. This represents a node mask, which sets the feature attributes of the target node to zero. Detailed Implementation
[0035] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.
[0036] like Figure 1 and Figure 2 As shown in the figure, this embodiment discloses a graph anomaly detection method based on low-rank contrastive learning and reconstruction, the specific details of which are as follows:
[0037] 1) Low-rank view generation and subgraph sampling: Obtain graph data and generate low-rank graph data by dimensionality reduction through Singular Value Decomposition (SVD). Then, use the restarted random walk algorithm to sample the subgraph to obtain the original view and the low-rank view. The detailed process is as follows:
[0038] 1.1) Obtaining Graph Data: The graph is represented in the following form: , including nodes ,in It is a node attribute matrix. It is the set of real numbers. It is the number of nodes. It is a data dimension. Representing the 1 node It is an adjacency matrix, in the adjacency matrix Chinese Represents a node and nodes Connection status, Represents a node and nodes There is a connecting edge between them. This indicates a node. and nodes There are no connecting edges between them, where ;
[0039] 1.2) SVD Decomposition: First, during the pre-training phase of the graph, random SVD decomposition is applied to the graph. and ,in and It corresponds to a matrix and The left singular matrix, and These are diagonal matrices, with their main diagonal elements representing and The singular values, and and It corresponds to a matrix and The right singular matrix is obtained; then, the first r singular values are retained to obtain the reconstructed node attributes. and adjacency matrix ,Right now and ,in and These represent the number of singular values, and yes and The low-rank approximation of , where and SVD decomposition is used to generate low-rank graph data, which is then used for the next step of subgraph sampling.
[0040] 1.3) Subgraph Sampling: The Restarted Random Walk (RWR) algorithm is used to extract subgraphs from both the original graph data and the low-rank graph data. For each node... Sample a size of The subgraphs are used to obtain the original views. and low-rank view ,in and It is the adjacency matrix and feature matrix formed by sampling subgraphs from the original graph data, and similarly... and It is the adjacency matrix and feature matrix formed by the sampled subgraphs of the low-rank graph data.
[0041] 2) Low-rank contrastive learning: Contrast pairs are constructed on the original view and the low-rank view for contrastive learning, and the contrastive learning loss is obtained. The detailed process is as follows:
[0042] 2.1) Constructing the contrast loss between views First, the target node Corresponding original view and low-rank view The inputs are fed into a single GCN layer to obtain the embeddings: and ,in It is a learnable weight matrix. and Represents the target node in two views In the characteristic matrix and The feature attributes are masked, i.e., their values are set to zero; then, the final representation of the two views is obtained using the average readout function within the GCN: and ,in Indicates the node number in the two views. and express The node with the sequence number is embedded and output by GCN. and These represent the subgraph embeddings of the original view and the low-rank view, respectively. This refers to the number of nodes in the corresponding view; then, the nodes are mapped to the same space where the views are embedded using MLP, i.e., the node mapping between the two views. and ,in and Representing the target node respectively In the properties of the original view and the low-rank view, It is an MLP mapping; then, the sigmoid function is used in the discriminator to obtain the target node. Similarity scores between the subgraph embeddings corresponding to the two views and the subgraph embeddings: and ,in It is the sigmoid function. These are the weight parameters corresponding to the discriminator; following the concept of contrastive learning, the similarity scores between the two views... and To request a more precise match, it can be expressed as: ,in Represented as the Frobenius paradigm;
[0043] 2.2) Constructing the in-view contrast loss First, construct the comparison pairs needed for contrastive learning, i.e., the target nodes. The sampled subgraphs and their sampling subgraphs constitute positive sample pairs, while the target node... With another node in the same batch The sampled subgraphs form negative sample pairs; then, the similarity scores of the positive and negative sample pairs are calculated separately in the low-rank view, i.e.: and ,in, For nodes Subgraph embedding of the low-rank view corresponding to the sampled subgraph Close to 1 Approaching 0, we then use BCE loss to train the contrast loss for the low-rank view, i.e., Similarly, a similarity score can also be obtained from the original view. and The contrast loss of the original view is trained using BCE loss, i.e. Next, the final node-subgraph contrast loss, i.e., the in-view contrast loss, is calculated: ,in These are parameters that control the weights of the original view and the lower-rank view;
[0044] 2.3) Integration yields contrastive learning loss By integrating the inter-view contrast loss and the intra-view contrast loss, and adding them together, we can obtain the contrastive learning loss. .
[0045] 3) Low-rank attribute reconstruction: Perform low-rank attribute reconstruction on the original view and the low-rank view to obtain the final reconstruction loss of the original view and the low-rank view, including the following steps:
[0046] 3.1) Calculate the reconstruction loss of the original view and the reconstruction loss of the low-rank view, i.e., the node... Reconstruction loss of the original view and nodes Reconstruction loss of the low-rank view ,in and They are and The series combination of adjacent nodes embedded in the middle It is an MLP mapping;
[0047] 3.2) Calculate the final reconstruction loss for the original view and the low-rank view: .
[0048] 4) Model Training: First, the final loss function is obtained by combining the reconstruction loss and the contrastive learning loss. ,in, These are the parameters that balance the reconstruction loss and the contrastive learning loss; finally, the loss function is used. Training a graph neural network model.
[0049] 5) Anomaly Score Detection: Based on the trained graph neural network model, anomaly detection is performed on nodes in the graph under test to identify potential anomaly nodes. This includes the following steps:
[0050] 5.1) Satisfying the form The graph to be tested is input into the trained graph neural network model, and steps 1) and 2) are repeated to obtain the corresponding representation, and then step 5.2) is executed.
[0051] 5.2) Calculate the anomaly score for low-rank view contrast learning. That is, the difference in similarity between negative sample pairs and positive sample pairs, expressed as: ,in The higher the score, the more abnormal the node is, and then the same abnormality score is obtained from the original view comparison learning. Finally, the outlier scores for the contrastive learning component were obtained. ;
[0052] 5.3) Calculate the reconstruction anomaly score for the low-rank view. ,in The higher the score, the more abnormal the node is, and the reconstruction anomaly score of the original view is obtained in the same way. Finally, the final anomaly score of the low-rank attribute reconstruction is obtained. ;
[0053] 5.4) Integrate the two abnormal scores above to obtain the final abnormal score of the node. The higher the score, the more abnormal the node, thus detecting potentially abnormal nodes.
[0054] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.
Claims
1. A graph anomaly detection method based on low-rank contrastive learning and reconstruction, characterized in that, Includes the following steps: 1) Low-rank view generation and subgraph sampling: Obtain graph data and generate low-rank graph data by dimensionality reduction through SVD singular value decomposition. Then, use the restarted random walk algorithm to sample the subgraph to obtain the original view and the low-rank view. 2) Low-rank contrastive learning: Contrast pairs are constructed on the original view and the low-rank view to perform contrastive learning and obtain contrastive learning loss; 3) Low-rank attribute reconstruction: Perform low-rank attribute reconstruction on the original view and the low-rank view to obtain the final reconstruction loss of the original view and the low-rank view; 4) Model training: Train the graph neural network model by combining contrastive learning loss and reconstruction loss; 5) Anomaly score discrimination: Based on the trained graph neural network model, anomaly discrimination is performed on the nodes in the graph under test to identify potential anomaly nodes.
2. The graph anomaly detection method based on low-rank contrastive learning and reconstruction according to claim 1, characterized in that, In step 1), the detailed process of low-rank view generation and subgraph sampling is as follows: 1.1) Obtaining Graph Data: The graph is represented in the following form: , including nodes ,in It is a node attribute matrix. It is the set of real numbers. It is the number of nodes. It is a data dimension. Representing the 1 node It is an adjacency matrix, in the adjacency matrix Chinese Represents a node and nodes Connection status, Represents a node and nodes There is a connecting edge between them. This indicates a node. and nodes There are no connecting edges between them, where ; 1.2) SVD Decomposition: First, during the pre-training phase of the graph, random SVD decomposition is applied to the graph. and ,in and It corresponds to a matrix and The left singular matrix, and These are diagonal matrices, with their main diagonal elements representing and The singular values, and and It corresponds to a matrix and The right singular matrix is obtained; then, the first r singular values are retained to obtain the reconstructed node attributes. and adjacency matrix ,Right now and ,in and These represent the number of singular values, and yes and The low-rank approximation of , where and SVD decomposition is used to generate low-rank graph data, which is then used for the next step of subgraph sampling. 1.3) Subgraph Sampling: The Restarted Random Walk (RWR) algorithm is used to extract subgraphs from both the original graph data and the low-rank graph data. For each node... Sample a size of The subgraphs are used to obtain the original views. and low-rank view ,in and It is the adjacency matrix and feature matrix formed by sampling subgraphs from the original graph data, and similarly... and It is the adjacency matrix and feature matrix formed by the sampled subgraphs of the low-rank graph data.
3. The graph anomaly detection method based on low-rank contrastive learning and reconstruction according to claim 2, characterized in that, In step 2), the detailed process of the low-rank contrastive learning is as follows: 2.1) Constructing the contrast loss between views First, the target node Corresponding original view and low-rank view The inputs are fed into a single GCN layer to obtain the embeddings: and ,in It is a learnable weight matrix. and Represents the target node in two views In the characteristic matrix and The feature attributes are masked, i.e., their values are set to zero; then, the final representation of the two views is obtained using the average readout function within the GCN: and ,in Indicates the node number in the two views. and express The node with the sequence number is embedded and output by GCN. and These represent the subgraph embeddings of the original view and the low-rank view, respectively. This refers to the number of nodes in the corresponding view; then, the nodes are mapped to the same space where the views are embedded using MLP, i.e., the node mapping between the two views. and ,in and Representing the target node respectively In the properties of the original view and the low-rank view, It is an MLP mapping; then, the sigmoid function is used in the discriminator to obtain the target node. Similarity scores between the subgraph embeddings corresponding to the two views and the subgraph embeddings: and ,in It is the sigmoid function. These are the weight parameters corresponding to the discriminator; following the concept of contrastive learning, the similarity scores between the two views... and To request a more precise match, it can be expressed as: ,in Represented as the Frobenius paradigm; 2.2) Constructing the in-view contrast loss First, construct the comparison pairs needed for contrastive learning, i.e., the target nodes. The sampled subgraphs and their sampling subgraphs constitute positive sample pairs, while the target node... With another node in the same batch The sampled subgraphs form negative sample pairs; then, the similarity scores of the positive and negative sample pairs are calculated separately in the low-rank view, i.e.: and ,in, For nodes Subgraph embedding of the low-rank view corresponding to the sampled subgraph Close to 1 Approaching 0, we then use BCE loss to train the contrast loss for the low-rank view, i.e., Similarly, a similarity score can also be obtained from the original view. and The contrast loss of the original view is trained using BCE loss, i.e. Next, the final node-subgraph contrast loss, i.e., the in-view contrast loss, is calculated: ,in These are parameters that control the weights of the original view and the lower-rank view; 2.3) Integration yields contrastive learning loss By integrating the inter-view contrast loss and the intra-view contrast loss, and adding them together, we can obtain the contrastive learning loss. .
4. The graph anomaly detection method based on low-rank contrastive learning and reconstruction according to claim 3, characterized in that, In step 3), the low-rank attribute reconstruction includes the following steps: 3.1) Calculate the reconstruction loss of the original view and the reconstruction loss of the low-rank view, i.e., the node... Reconstruction loss of the original view and nodes Reconstruction loss of the low-rank view ,in and They are and The series combination of adjacent nodes embedded in the middle It is an MLP mapping; 3.2) Calculate the final reconstruction loss for the original view and the low-rank view: .
5. The graph anomaly detection method based on low-rank contrastive learning and reconstruction according to claim 4, characterized in that, In step 4), the model training includes: First, by combining the reconstruction loss and the contrastive learning loss, the final loss function is obtained. ,in, These are the parameters that balance the reconstruction loss and the contrastive learning loss; finally, the loss function is used. Training a graph neural network model.
6. The graph anomaly detection method based on low-rank contrastive learning and reconstruction according to claim 5, characterized in that, In step 5), the anomaly score determination includes the following steps: 5.1) Satisfying the form The graph to be tested is input into the trained graph neural network model, and steps 1) and 2) are repeated to obtain the corresponding representation, and then step 5.2) is executed. 5.2) Calculate the anomaly score for low-rank view contrast learning. That is, the difference in similarity between negative sample pairs and positive sample pairs, expressed as: ,in The higher the score, the more abnormal the node is, and then the same abnormality score is obtained from the original view comparison learning. Finally, the outlier scores for the contrastive learning component were obtained. ; 5.3) Calculate the reconstruction anomaly score for the low-rank view. ,in The higher the score, the more abnormal the node is, and the reconstruction anomaly score of the original view is obtained in the same way. Finally, the final anomaly score of the low-rank attribute reconstruction is obtained. ; 5.4) Integrate the two abnormal scores above to obtain the final abnormal score of the node. The higher the score, the more abnormal the node, thus detecting potentially abnormal nodes.