Semi-supervised Classification of Fraud Data in Commercial Auctions

Given the magnitude of monetary transactions at auction sites, they are very attractive to fraudsters and scam artists. Shill bidding (SB) is a severe fraud in e-auctions, which occurs during the bidding period and is driven by modern-day technology and clever scammers. SB does not produce any obvious evidence, and it is often unnoticed by the victims. The lack of availability of training datasets for SB and the difficulty in identifying the behavior of sophisticated fraudsters hinder research on SB detection. To safeguard consumers from dishonest bidders, we were incentivized to investigate semi-supervised classification (SSC) for the first time, which is the most suitable approach to solving fraud classification problems. In this study, we first introduce two new SB patterns, and then based on a total of nine SB patterns, we build an SB dataset from commercial auctions and bidder history data. SSC requires the labeling of a few SB data samples, and to this end, we propose an anomaly detection method based on data clustering. We addressed the skewed class distribution with a hybrid data sampling method. Our experiments in training several SSC models show that using primarily unlabeled SB data with a few labeled SB data improves predictive performance when compared to that of supervised models.