Research On Anomaly Detection Methods Of Relational Networks Based On Graph Embedding

As an important information carrier,relational networks are able to model and express complex information about the associations between different types of data objects.With the development of the Internet,the data size of relational networks is increasing.How to learn and apply relational networks effectively is a hot research topic in the field of machine learning and data mining in recent years.Graph embedding methods can effectively handle relational network data and capture the complex relationships between nodes and edges in relational networks,and can be applied to various relational network anomaly detection tasks such as transaction networks,social networks and the Internet of Things.However,existing graph embedding-based relational network anomaly detection methods have certain shortcomings.Among them,static graph embedding-based relational network anomaly detection methods often ignore strong structural similarity,while dynamic graph embedding-based relational network anomaly detection methods have poor application capabilities and often ignore the differences in temporal information between different subgraph nodes.To address these problems,this paper focuses on the detection of anomalous targets in relational networks from two perspectives: static graph embedding methods and dynamic graph embedding methods:(1)Static graph embedding method based on multi-order weighted walking and its application researchTo address the existing static graph embedding-based relational network anomaly detection methods that ignore strong structural similarity,this paper proposes a static graph embedding method based on multi-order weighted walking.The method uses a node similarity evaluation strategy to learn the structural similarity and attribute similarity of nodes,and uses multi-order weighted walking to obtain a sequence of similar nodes,and then aggregates similar nodes at a long distance to extract the embedding features of nodes.In the transaction network fraud user detection,based on the embedding features extracted by this static graph embedding method,a logistic regression algorithm is used to classify users in the transaction network and identify fraudulent users in the transaction network.The experimental results show that the method has better anomaly detection performance and is effective in detecting fraudulent users in the transaction network.(2)Dynamic graph embedding method based on spatio-temporal coding and adversarial autoencoder and its application researchThis paper proposes a dynamic graph embedding method based on spatio-temporal coding and adversarial autoencoder to address the poor application capability of existing dynamic graph embedding-based relational network anomaly detection methods and the neglect of temporal information differences between different sub-graph nodes.The method uses three encoding methods to obtain the spatio-temporal encoding information of the target edges,and uses a semi-supervised adversarial autoencoder to reconstruct the target edges and extract the embedding features of the target edges.In identifying anomalous interactions in the relational network,the method discriminates whether the target edge is anomalous by comparing the difference between the target edge and the reconstructed edge embedding features.And only normal data is required to train the adversarial autoencoder model,which can effectively improve its application in different relational network anomaly detection scenarios.Experimental results on multiple benchmark relational network datasets show a significant improvement in the performance of the method in terms of anomalous interaction detection performance.Finally,a comprehensive summary of the paper’s research is presented and future research directions are envisaged.