SPAN: Finding collaborative frauds in online auctions

Fraud is an ongoing concern for online auction websites. Current methods to detect or prevent fraud have been limited in several ways, making them difficult to apply in real world settings. Firstly, existing methods cannot adapt to changes in the behaviour of fraudulent users over time: new models must be continuously constructed as they gradually lose accuracy. In addition, each method can only be used to detect a specific type of fraud. Secondly, existing methods are generally poor at identifying collaborative frauds. And thirdly, method training and evaluation has been limited by the quality of available datasets. We propose an algorithm named SPAN (Score Propagation over an Auction Network), for detecting users committing collaborative fraud that addresses these problems. SPAN is a two phase method that first applies anomaly detection on multiple 2-dimensional feature subspaces to generate an initial anomaly score for each user, then applies belief propagation to revise those scores to identify suspicious groups of users. We report extensive experimental results using synthetic data which shows that SPAN performs well across three different types of fraud, and outperforms a previous approach for collaborative fraud detection called 2-Level Fraud Spotting. Experiments on a real dataset shows that SPAN behaves reasonably, and can identify groups of users that appear anomalous.