EntroBot: Entropy-Based Modeling of Account Behavior for Social Bot Detection

Abstract

Online social platforms have transformed information dissemination, but their openness has facilitated the rise of social bots that actively influence public discourse at scale. Among existing approaches to social bot detection, behavior-based methods are recognized for robustness and interpretability, though their restricted temporal scope hinders effective modeling of behavioral dynamics and constrains detection accuracy. In this study, we introduce EntroBot, an entropy-based method that detects social bots by modeling account behavior across multiple timescales. Empirical analysis of entropy-quantified behavioral patterns reveals that genuine users maintain behavioral regularity across extended time scales, with stability increasing relative to social bots. Experiments show that EntroBot outperforms existing behavior-driven approaches in detection accuracy and rivals deep learning methods, while preserving interpretability and robustness. Ablation analysis shows that single-scale features are insufficient to characterize diverse bot behaviors, but integrating short-term, long-term, and dynamic signals yields synergistic gains.