Thinking Like a Fraudster: Detecting Fraudulent Transactions via Statistical Sequential Features

Abstract

Aiming at the increasing threat of fraud in electronic transactions, so far researchers have already proposed many different models. However, few previous studies take advantage of the sequential characteristics of fraudulent transactions. In this paper, by statistical analysis on a real dataset, we discover that partial-order sequential features are able to reflect the intrinsic motivation of fraudsters, e.g., stealing the money as quickly as possible before being intercepted. Based on the sequential features, we propose a novel model, SeqFD (Sequential feature boosting Fraud Detector), to detect fraudulent transactions real-timely. SeqFD applies a sliding time window strategy to aggregate the historical transactions. In specific, statistical sequential features are computed based on the transactions within the time window. Thus, the raw dataset can be transformed into a feature set. Several classification models are evaluated on the feature set, and finally, XGBoost is validated to be a fast, accurate and robust classifier which fits well with SeqFD. The experiments on real dataset show that the proposed model reaches a 97.2% TPR (True Positive Rate) when FPR (False Positive Rate) is less than 1%. Furthermore, the average time for giving a prediction is 1.5 ms, which meets the real-time requirement in the industry.

This work was supported in part by the National Natural Science Foundation of China under Grant 61332008, in part by the National Key Research and Development Program of China under Grant 2018YFC0831403.