A Supervised Learning Process to Elicit Fraud Cases in Online Auction Sites

摘要

Fraud is a recurring phenomenon at online actions sites like eBay. The enormous amount of transaction data public ally available offers a good opportunity for fraud prevention based on learning methods. However, online auction sites usually neither confirm nor deny fraudulent behavior: they simply suspend seller accounts and publicize feedback information supplied by buyers. While some cases receive media attention, most of them are hidden in the site's database. This limits the possibility of developing and testing new learning methods for fraud prevention, due to the scarcity of fraud samples. In order to overcome this limitation, we designed a system based on supervised learning to recognize in the textual comments left by buyers some common statements regarding seller behavior. Combining the type and frequency of those statements with other public ally available data, we can build a set of sellers who can arguably be considered fraudsters. We implemented a prototype of the system and evaluated it using data extracted from a major online auction site.