Network-based spam filter on Twitter

Rapidly growing micro-blogging social networks, such as Twitter, have been infiltrated by large number of spam accounts. Limited to 140 characters, Twitter spam is often vastly different from traditional email spam and link spam such that conventional methods of content-based spam filtering are insufficient. Many researchers have proposed schemes to detect spammers on Twitter. Most of these schemes are based on the features of the user account or the features in the content, such as the similarity or the ratio of URLs. In this paper, we propose a network analysis based spam filter for Twitter. By analyzing the network structure and relations between senders and receivers, this spam filter does not require large data collection up front, thus is able to provide almost real-time detection for spams. By using the public API methods provided by Twitter, our system crawled the users in the sub-graph between suspicious senders and receivers. Then we analyze the structure and the properties of the sub-graph and compare them with those we collected from legitimate senders and receivers. Our study showed that spammers rarely have a network distance of 4 or less from their victim. Using a sub-graph of diameter 4 constructed between sender and receiver, we can further improve the recall of our spam filter with promising results by utilizing network-based features such as number of independent paths and normalized page ranks.