Research On Interpretable Financial User Anomaly Transaction Detection

With the rise of Internet finance,more and more companies are entering the financial industry.Although the financial boom of the Internet has added vitality to the industry,it has also brought new challenges to both Internet companies and regulatory authorities.For Internet companies,with the emergence of more and more new cheating behaviors such as "marketing cheating" and "shilling attack",traditional methods for bank card fraud detection can hardly meet the complex financial risk control scenarios’ demands nowadays.However,although the existing anomaly detection models based on deep learning and graph learning perform well,they usually fail to provide convincing explanations for anomalies,which is neither conducive to the improvement of the risk control system,nor to the supervision of relevant departments.With the development of explainable machine learning,we can make anomaly detection systems more intelligent,so that they can identify and explain anomalies.Therefore,how to design a user and transaction anomaly detection system with high accuracy and high interpretability is an important problem that needs to be solved in the field of financial supervision.In view of the above-mentioned industry needs and the shortcomings of existing research,this paper designs an anomaly detection model for bank personal account fund sequence scenarios and an anomaly detection model for transaction graph scenarios.Then,interpreter models are designed for the above two models to explain anomalies.The specific work of this paper is as follows:(1)Data set construction.Our research uses real bank data and real Bitcoin transaction network data to construct a standard users’ funding sequence data set and a Bitcoin transaction graph data set,which provides a data basis for the subsequent design of sequence-level and graph-level anomaly detection systems.(2)Anomaly detection and interpretable model design in the scenario of users’ funding sequence.For the bank’s personal account funding sequence,the data presents time series.In our research,continuous funding sequences are quantified into buckets,and it is the first time to introduce sequence graph transformation to extract sequence embedding features.Next,we build a deep neural network to train a sequence classifier model.Then,the interpreter model is designed using the local interpretability method to achieve the effect of black-box interpretation of specific outliers,and the interpretation results of the white-box model are compared.The experimental results show that the anomaly detection model designed in this study has both high accuracy and high interpretability.(3)Anomaly detection and interpretable model design in transaction graph scenario.Our research draws on the design ideas of the perturbation-based method and the surrogate method,and specially designs a transaction graph anomaly detection and interpretability model that takes both the influence of node characteristics and neighborhoods into account.In addition,our study adopts a new mask generation algorithm,using a single model to simultaneously explain the importance of node features and neighboring nodes,and proposes the "noise ratio" for the first time to evaluate the model’s anti-noise performance.Experiments show that our model outperforms existing methods in terms of accuracy,interpretability and noise immunity.(4)Prototype system design.Based on the above research work,this research deploys the anomaly detection and interpretable models of the above two scenarios,and designs and implements a real-time online interpretable user abnormal transaction detection prototype system.The system can call the corresponding interpreter model according to the specific abnormal point input by the user,and visualize the contribution of each feature to the classification result.