A Semi-Supervised Framework for Social Spammer Detection

Spammers create large number of compromised or fake accounts to disseminate harmful information in social networks like Twitter. Identifying social spammers has become a challenging problem. Most of existing algorithms for social spammer detection are based on supervised learning, which needs a large amount of labeled data for training. However, labeling sufficient training set costs too much resources, which makes supervised learning impractical for social spammer detection. In this paper, we propose a semi-supervised framework for social spammer detection(SSSD), which combines the supervised classification model with a ranking scheme on the social graph. First, we train an original classifier with a small number of labeled data. Second, we propose a ranking model to propagate trust and distrust on the social graph. Third, we select confident users that are judged by the classifier and ranking scores as new training data and retrain the classifier. We repeat the all steps above until the classifier cannot be refined any more. Experimental results show that our framework can effectively detect social spammers in the condition of lacking sufficient labeled data.