Detecting Spammers with Changing Strategies via a Transfer Distance Learning Method

Social spammers bring plenty of harmful influence to the social networking involving both social network sites and normal users. It is a consensus to detect and filter spammers. Existing social spammer detection approaches mainly focus on discovering discriminative features and organizing these features in a proper way to improve the detection performance, e.g., combining multiple features together. However, spammers are easy to escape being detected by using changing spamming strategies. Various spamming strategies bring differences in data distribution between training and testing data. Thus, previous fixed approaches are difficult to achieve desired performance in real applications. To address this, in this paper, we present a transfer distance learning approach, which combines distance learning and transfer learning to extract informative knowledge underlying training and testing instances in a unified framework. The proposed approach is validated on large real-world data. Empirical experiments results give the evidence that our method is efficient to detect spammers with changing spamming strategies.