Online auction has been proposed as a trading platform on internet for more than a decade. It attracts not only legitimate users who attempt to sell their products but also fraudulent users, who desire to commit false transactions for deceiving the third party. It is difficult to separate fraudsters from legitimate users only relying on reputation scores of the traders. Accordingly, fraudster detection is important to ensure the continued success of online auctions. The purpose of this dissertation is to increase the classification performance for detecting fraudsters with inflated reputation. We propose a novel framework for fraudster detection based on concept of neighbor features. The neighbor features of an auction account are calculated from the feature of all traders that have transactions with this account.
In order to distinguish fraudsters from non-fraudsters, we propose a framework that differentiates fraudsters from non-fraudsters based on various features of each trader (i.e., various types of neighbor diversity and various neighbor-driven attributes). The proposed framework consists of three parts, i.e., neighbor diversity based on Shannon entropy, various forms of neighbor diversity, and neighbor-driven attribute. In this dissertation, various forms of neighbor features have been calculated using different approaches. Several features of neighbor have been used for detecting fraudsters from online auction transactions. Real world online auction dataset crawled from Ruten has been used for conducting the experiments in this dissertation. The results indicate that neighbor features enhance overall classification performance, compared to the state-of-the-art methods that use k-core and center weight.