Paper detail

Online learning for Social Spammer Detection on Twitter

Social networking services like Twitter have been playing an import role in people's daily life since it supports new ways of communicating effectively and sharing information. The advantages of these social network services enable them rapidly growing. However, the rise of social network services is leading to the increase of unwanted, disruptive information from spammers, malware discriminators, and other content polluters. Negative effects of social spammers do not only annoy users, but also lead to financial loss and privacy issues. There are two main challenges of spammer detection on Twitter. Firstly, the data of social network scale with a huge volume of streaming social data. Secondly, spammers continually change their spamming strategy such as changing content patterns or trying to gain social influence, disguise themselves as far as possible. With those challenges, it is hard to directly apply traditional batch learning methods to quickly adapt newly spamming pattern in the high-volume and real-time social media data. We need an anti-spammer system to be able to adjust the learning model when getting a label feedback. Moreover, the data on social media may be unbounded. Then, the system must allow update efficiency model in both computation and memory requirements. Online learning is an ideal solution for this problem. These methods incrementally adapt the learning model with every single feedback and adjust to the changing patterns of spammers overtime. Our experiments demonstrate that an anti-spam system based on online learning approach is efficient in fast changing of spammers comparing with batch learning methods. We also attempt to find the optimal online learning method and study the effectiveness of various feature sets on these online learning methods.