Research And Application Of Fraud Detection Method Based On Graph Features

With the continuous integration and development of the Internet and finance,various products such as Internet finance and mobile payment have been derived,tightening the connection between people and finance.At the same time,a large number of financial frauds have also occurred,causing huge economic losses for individuals and financial institutions.Online fraud detection technology can quickly and effectively prevent fraudulent transactions and reduce losses.For practical applications,fraud detection needs to be real-time,high efficient,stable and interpretable.To meet the requirements above,so far,the industry mainly uses stream processing technology to process historical transaction data for cumulative features,and then uses a classification algorithm to train to obtain a detection model.With the continuous development of graph theory,graph algorithms play an increasingly important role in the field of fraud detection.This thesis first introduces the relevant background,the significance of the arguments,and the current research status in this field.After that,the related technologies such as graph database,community detection algorithm,Spark distributed computing framework and random forest algorithm are elaborated.In this thesis,Neo4j graph database is used to obtain graph structure features in real time based on original transaction dataset.The two algorithms of Louvain and BMLPA implemented by graph parallel computing technology are used to obtain community features.By introducing the graph structure features and the community features into the classification algorithm,the detection ability of the model is improved.The main contributions of this thesis are:1)Proposing a methodology for introducing graph structure features and community features based on common features of industry.Through a large number of experiments on real transaction dataset,it is verified that this method significantly improves the fraud detection ability of the model;2)Implementing the parallelization of community detection algorithms using GraphX distributed graph computing framework,which can be applied to transaction datasets to obtain community features;3)Designing and implementing a method of extracting graph structure features in real time based on graph database,which can be easily combined with existing methods in industry.