Using Self-Organizing Maps for Fraud Prediction at Online Auction Sites

Abstract

Online auction sites have to deal with a enormous amount of product listings, of which a fraction is fraudulent. Although small in proportion, fraudulent listings are costly for site operators, buyers and legitimate sellers. Fraud prediction in this scenario can benefit significantly from machine learning techniques, although interpretability of model predictions is a concern. In this work we extend an unsupervised learning technique -- Self-Organizing Maps -- to use labeled data for binary classification under a constraint on the proportion of false positives. The resulting technique was applied to the prediction of non-delivery fraud, achieving good results while being easier to interpret.