| With the development of information technology,graph,as a convenient and effective modeling method,is widely used to represent complex structured data.Graph anomaly detection has a wide range of applications in real life,such as malicious account detection in social networks and fraud detection in financial networks.The data of the graph anomaly detection has the characteristics of complex topological structure relationship,diverse information,and unbalanced labels,which causes the performance of the existing anomaly detection algorithms to be unsatisfactory and affects the performance of anomaly detection tasks.In order to effectively deal with the problem of anomaly detection and improve the performance of anomaly node detection,the main contributions of this paper are as follows:（1）To solve the problems caused by the lack of target in homogeneous anomaly node detection techniques and the problems caused by the mixing of task-independent noise edges into the neighborhood aggregation,we propose an homogeneous anomaly detection method that incorporating network structure with node information on attributed graphs.This technology first designs a graph neural networkbased semi-supervised loss function for anomaly detection tasks to detect structural anomalies;At the same time,a symmetric deep neural network is constructed to detect attribute anomalies and alleviate noisy edges in the process of aggregation.In addition,experiments on five datasets verify the effectiveness of the model.（2）In view of how to extract the unique information of each view,learn consistent information between views,and measure the importance of each views for multiview graph representation learning techniques,an unsupervised and contrastive learning based heterogeneous graph representation learning technique is designed.First,the model uses the contrastive learning to help the encoder extract discriminative node features;Then,the model uses the contrastive learning technique to learn the consistency between different views;Finally,a mutual information-based view importance measurement method is proposed to generate the final node representation.We validate the effectiveness of the learned node representations on graph analysis tasks such as node classification,clustering,similarity search,and fraud detection on multiple datasets.（3）In view of the camouflage problem,data sparse problem and class imbalance problem in the fraud detection scene of heterogeneous graph,an anomaly node detection framework based on enhanced graph learning is proposed.In this framework,a parameterized neural network is used to obtain the similarity between nodes to alleviate the problem of feature camouflage;Then,the similarity is filter noise neighbors and add neighbors to increase the correlation of similar nodes and reduce the correlation of dissimilar nodes,thereby alleviating the problem of relationship camouflage and data sparsity;At the same time,the framework enhances the quality of node representation through mutual information learning;Finally,considering the class imbalance problem in real scenarios,a class-aware model is designed.The loss function further improves the anomaly detection effect of the model.Experiments on several public datasets validate the effectiveness of framework,and ablation experiments and parameter experiments also verify the role of each module. |
