Feature Engineering for Anti-Fraud Models Based on Anomaly Detection

Abstract

The paper presents two algorithms as a solution to the problem of identifying
fraud intentions of a customer. Their purpose is to generate variables that
contribute to fraud models’ predictive power improvement. In this article,
a novel approach to the feature engineering, based on anomaly detection, is
presented. As the choice of statistical model used in the research improves
predictive capabilities of a solution to some extent, most of the attention
should be paid to the choice of proper predictors. The main finding of the
research is that model enrichment with additional predictors leads to the further
improvement of predictive power and better interpretability of anti-fraud model.
The paper is a contribution to the fraud prediction problem but the method
presented may generate variable input to every tool equipped with variableselection algorithm. The cost is the increased complexity of the models obtained.
The approach is illustrated on a dataset from one of the European banks.