Research On Self-supervised Heterogeneous Graph Representation Learning And Its Application

In recent years,the vigorous development of internet finance has accelerated the process of financial digitization,bringing people a more convenient life,but also attracting a large number of fraudsters.Financial fraud accompanies transactions between users and naturally forms a user interaction graph.In order to effectively detect fraudsters,academia and industry model user interaction in financial scenarios as a heterogeneous graph,and use heterogeneous graph representation learning to extract features of heterogeneous graphs,revealing the suspiciousness of these users at the graph level.In this thesis,self-supervised heterogeneous graph representation learning and its application in financial fraud detection is studied.Targeting the problems of insufficient labeled data in the application of heterogeneous graph representation learning and the difficulty of detecting camouflaged fraudsters,the shortcomings of existing work are analyzed.This thesis proposes improved methods for two difficult problems.The main contributions of the thesis are as follows:1.Heterogeneous graph representation learning faces the problem of insufficient labeled data in their applications.To address this problem,a method for heterogeneous graph neural networks representation learning based on contrastive learning is proposed.By cross-contrasting multiple views and maximizing the consistency between them,the correlations between different meta-paths are captured.Attention mechanism is used to fuse different meta-paths and maximize the mutual information between the fused nodes and graph summary to exploit high-order semantic information.Furthermore,a data augmentation strategy based on multi-view neighborhood connections is adopted to reduce the negative effects of irrelevant nodes.Experimental results show that on the classification task of the DBLP dataset,the proposed method has a 7.5% improvement in the Mi F1 index in comparison with the graph convolutional network.2.Fraudsters hide among the normal user community by using disguised transactions,so that their characteristics are assimilated by surrounding benign entities and make them difficult to detect.To address this problem,a fraudsters detection method based on heterogeneous graph neural network is proposed.In this method,the community discovery algorithm is applied to the entity relationship graph generated by transaction users and devices to realize the separation of fraudster groups and normal user groups.By applying the pruning algorithm,the hidden camouflage of fraudsters is removed to reduce the impact on the detector.The further application of the heterogeneous graph neural network improves the ability to mine hidden fraudsters.Experimental results show that the proposed method has a 27.54% improvement in the AUCPR index compared to the graph convolutional network on the given dataset.3.By combining the two algorithms proposed above,a financial fraud detection system is designed and implemented,which is suitable for camouflaged fraudsters detection under insufficient labeled data.Users can upload data,configure models,and predict potentially fraudsters on the front-end visual interface to complete fraud detection conveniently and easily.