A Two-Stage Risk Model Building and Evaluation in Reject Inference

Abstract: With the prevalence of the electronic commerce, fraud commonly exists in online business. To prevent fraud transactions, suspicious transactions are rejected with unknown fraud statuses by the online decision system. To build an accurate risk model for fraud detection, one great challenge is how to use the information of rejected transactions for model construction and model evaluation. In this work, we propose a two-stage model to effectively identify known fraud patterns and missing frauds. The proposed model fully exploits the fraud patterns in both accepted and rejected transactions, thus enhances the detection of fraud at both stages. Moreover, we develop a novel model evaluation criterion based on the adjusted net profit value, which takes the information of rejected transactions into account. It therefore can be used for efficiently finding the optimal risk model. The performance of the proposed method is illustrated through a real case study of Microsoft Xbox online transaction data.

Bio: Xinwei Deng is an assistant professor at department of statistics in Virginia Tech since 2011. He got his Ph.D from Georgia Tech in 2009, and spent two years at University of Wisconsin-Madison as visiting assistant professor from 2009-2011. His research interests are Interface between design of experiments and machine learning, model and analysis of high-dimensional data, and Covariance matrix estimation and its applications.