Online Social Spammer Detection Based on Semi-supervised Learning

Abstract

Social networks, like Twitter and Sina Weibo have become popular platforms for internet users to disseminate and share information. However, there is a massive number of spammers who have been conducting social spamming on normal users. Social spammers have severely influenced the user experience and the healthy development of the website. Therefore, detecting spammers on social networks has become an urgent problem. Most of the existing methods for spammer detection is based on supervised learning, which is inappropriate for implementing in practice, as labeling sufficient training data costs too much resources. In this paper, we propose a semi-supervised solution for effective spammer detection which combines co-training with social graph model. First, we utilize the behavior features and content features to train two original classifiers. Then we select confident users judged by their credibility scores which is based on the social graph model. Afterwards, these users selected are added to the training set to retrain the two classifiers. The two steps are repeated till the classifiers cannot be refined any more. Experimental results prove that this method is capable of detecting social spammers effectively without sufficient labeled data.