Practical Detection of Click Spams Using Efficient Classification-Based Algorithms

Abstract

Most of today’s Internet services utilize user feedback (e.g. clicks) to improve the quality of their services. For example, search engines use click information as a key factor in document ranking. As a result, some websites cheat to get a higher rank by fraudulently absorbing clicks to their pages. This phenomenon, known as “Click Spam”, is initiated by programs called “Click Bot”. The problem of distinguishing bot-generated traffic from the user traffic is critical for the viability of Internet services, like search engines. In this paper, we propose a novel classification-based system to effectively identify fraudulent clicks in a practical manner. We first model user sessions with three different levels of features, i.e. session-based, user-based and IP-based features. Then, we classify sessions with two different methods: a one-class and a two-class classification that both work based on the well-known K-Nearest Neighbor algorithm. Finally, we analyze our methods with the real log of a Persian search engine. Experimental results show that the proposed algorithms can detect fraudulent clicks with a precision of up to 96% which outperform the previous works by more than 5%.

Keywords

• bot
• click spam
• user session modeling
• classification