Ecommerce Fraud Detection Through Fraud Islands and Multi-layer Machine Learning Model

Abstract

Main challenge for e-commerce transaction fraud prevention is that fraud patterns are rather dynamic and diverse. This paper introduces two innovative methods, fraud islands (link analysis) and multi-layer machine learning model, which can effectively tackle the challenge of detecting diverse fraud patterns. Fraud Islands are formed using link analysis to investigate the relationships between different fraudulent entities and to uncover the hidden complex fraud patterns through the formed network. Multi-layer model is used to deal with the largely diverse nature of fraud patterns. Currently, the fraud labels are determined through different channels which are banks’ declination decision, manual review agents’ rejection decisions, banks’ fraud alert and customers’ chargeback requests. It can be reasonably assumed that different fraud patterns could be caught though different fraud risk prevention forces (i.e. bank, manual review team and fraud machine learning model). The experiments showed that by integrating few different machine learning models which were trained using different types of fraud labels, the accuracy of fraud decisions can be significantly improved.