Machine Learning Algorithms for Detecting and Analyzing Social Bots Using a Novel Dataset

Abstract

Social media is internet-based technology and an electronic form of communication that facilitates sharing of ideas, documents, and personal information. Twitter is a microblogging platform and is the most effective social service for posting microblogs and likings, commenting, sharing, and communicating with others. The problem we are shedding light on in this paper is the misuse of bots on Twitter. The purpose of bots is to automate specific repetitive tasks instead of human interaction. However, bots are misused to influence people’s minds by spreading rumors and conspiracy related to controversial topics. In this paper, we initiate a new benchmark created on a 1.5M Twitter profile. We train different supervised machine learning on our benchmark to detect bots on Twitter. In addition to increasing benchmark scalability, various autofeature selections are utilized to identify the most influential features and remove the less influential ones. Furthermore, over-under-sampling is applied to reduce the imbalance effect on the benchmark. Finally, our benchmark compared with other stateof-the-art benchmarks and achieved a 6% higher area under the curve than other datasets in the case of generalization, improving the model performance by at least 2% by applying over-/undersampling.