Research On Edge Anomaly Detection Technology Of Graph

With the development of Internet technology,people's lives are related to the online world all the time,and a large amount of data is also generated in the process.Due to the existence of some malicious services or attack methods,there are usually abnormal behaviors in these data,which brings huge security risks and economic losses to people,such as false attention in social friend data and fraudulent transactions in electronic payment data.A common solution is to use anomaly detection methods to identify abnormal behaviors.Traditional anomaly detection methods mainly rely on non-graph data structures and do not consider the topological features of graphs,so the detection model can easily be bypassed by forged features.The topological features of graph data are usually not easy to be forged,so abstracting the data into a graph data structure can further improve the anomaly detection effect,such as abstracting social friend data into a social network.Therefore,based on the graph data structure,this thesis studies the edge anomaly detection of static graphs and the edge anomaly detection of dynamic graphs.The research work and innovation are as follows:1)In the static graph,for the problem of anomaly detection of spurious edges different from real edges,the existing methods regard the data containing abnormal samples as the data of all normal samples to learn the model.That is,the characteristics of abnormal samples cannot be directly learned,and the influence of abnormal samples on learning cannot be considered.Therefore,a LATA model based on contrastive learning is proposed.Its link information augmentation module performs sample enhancement based on network structure distribution sampling,which can reduce network structure changes and generate abnormal samples? its link-level contrastive denoising module imposes consistency and difference constraints on node vectors of different views and attenuates the weight of reconstruction loss,which can force the model to learn the characteristics of abnormal edges and reduce the influence of abnormal samples? its link information correction module corrects samples with large loss distribution as abnormal samples,which can correct abnormal edges with a high hit rate and improve the accuracy of network reconstruction.Experiments and analyses are conducted on six real datasets,which confirm that the proposed model is more effective and robust than baseline methods in detecting abnormal edges.2)Aiming at the problem of anomaly detection of fraudulent transaction edges in transaction networks in dynamic graphs,existing methods use GNN to extract graph structure features and input RNN to extract time series features.The graph structure is separated from the time series feature extraction,and the label information is used for node embedding and the impact of long-tail nodes on embedding is not considered.Therefore,an ETGE model based on time series features is proposed to detect abnormal edges.Its feature extractor module extracts single transaction features and common transaction features in the ego-net network,which can extract The label information of historical transaction edges is directly used for embedding vector generation? its missing information prediction module learns the relationship model through the change process of the missing neighbor information sequence,which can predict the missing neighbor information of the node and supplement the aggregation? its evolution module uses the MGGRU unit to learn the node sequence change information from the forward edge and reverse edge direction,which can capture the network structure features while extracting the node timing features.Finally,experiments and analyses are conducted on three real datasets,which confirm that the proposed model performs better than the baseline method in detecting anomalous edges.