Social Bot Detection Through Model-Based Time Series Clustering

Abstract

In order to maintain the health and safety of online communities, it is important to understand and detect potential threats. Because of the possibility of large-scale impacts achieved through automation, social bots can manipulate and disrupt the experiences of users on social networking applications. When a group of bots coordinate, such disruptions can become systemic. Because members of these botnets act together, they have the potential to be detected through clustering of temporal features, even in the absence of existing labeled data. This work investigates how event sequences from social media applications may be used in a model-based clustering framework for the detection of groups of bots. The use of model-based clustering should allow for the use of formal model selection techniques, avoiding the problem of hyper-parameter optimization while aiding in the explanation of what constitutes suspicious behavior. Based on applications to both simulated and real-world data, this work provides evidence against the primary recommended model selection technique while affirming the utility of model-based clustering of event sequences in detection and classification schemes.