Using Self-labeling and Co-Training to Enhance Bots Labeling in Twitter

The rapid evolution in social bots have required efficient solutions to detect them in real-time. In fact, obtaining labeled stream datasets that contains variety of bots is essential for this classification task. Despite that, it is one of the challenging issues for this problem. Accordingly, finding appropriate techniques to label unlabeled data is vital to enhance bot detection. In this paper, we investigate two labeling techniques for semi-supervised learning to evaluate different performances for bot detection. We examine self-training and co-training. Our results show that self-training with maximum confidence outperformed by achieving a score of 0.856 for F1 measure and 0.84 for AUC. Random Forest classifier in both techniques performed better compared to other classifiers. In co-training, using single view approach with random forest classifier using less features achieved slightly better compared to single view with more features. Using multi-view of features in co-training in general achieved similar results for different splits.