Identification of Spambots and Fake Followers on Social Network via Interpretable AI-Based Machine Learning

Social networking platforms like X (Twitter) serve as hubs for open human interaction, but they are also increasingly infiltrated by automated accounts masquerading as human users. These bots often engage in activities such as spreading fake news and manipulating public opinion during politically sensitive times like elections. Most of the current bot detection methods rely on black-box algorithms, raising concerns about their transparency and practical usability. This study aims to address these limitations by developing a novel methodology for the detection of spambots and fake followers using annotated data. To this end, we propose an interpretable machine learning (ML) framework, leveraging multiple ML algorithms with hyperparameters optimized through cross-validation, to enhance the detection process. Furthermore, we analyze several features and provide a unique feature set that is optimized to offer excellent performance for bot detection. Moreover, we utilize multiple interpretable AI techniques which include Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). SHAP will help to display the effects of particular characteristics on the model’s prediction which will help in determining whether an account is a bot or a legitimate user. LIME will help to comprehend the model’s predictions, offering clarity regarding the traits or attributes that drive the classification conclusion. LIME allows researchers to detect bot-like activity in social networks by generating locally faithful explanations for each prediction. Our model offers enhanced interpretability by clearly highlighting the impact of various features used for spam and fake follower detection when compared to existing state-of-the-art social network bot detection systems. The results showcase the model’s ability to identify key distinguishing attributes between bots and legitimate users which offers a transparent and effective solution for social network bot detection. Additionally, we utilize two comprehensive datasets including Cresci-15 and Cresci-17, which serve as robust baselines for comparison. Our model showcases its effectiveness by outperforming other methods while providing interpretability which increases performance and reliability for the task of bot detection.