A novel integrated framework based on multi-view features for multidimensional social bot detection

Abstract

The ravage of malicious bots in online social networks has profoundly affected normal users. In this article, we propose a novel integrated framework to detect social bots. Specifically, social bot detection is performed from two dimensions: binary-class and fine-grained detection. Moreover, 35 features from three views are extracted to detect social bots, including eight newly defined features. Then, a category balancing based on resampling technology is designed to balance the training data. Finally, a divide-and-conquer strategy is integrated into Random Forest, and the interference of noise in the training process is reduced. Feature effectiveness evaluation found that extracting features from multi-views can describe bots more comprehensively. It is also noted from the category imbalance test that the balanced data set can prevent the detection result from tilting. Comparative experiments show that the integrated framework is more effective than the baseline both in social bot detection and the type detection of bots.