VeriBot: Evidence-Routed Social Bot Detection with Auditable Decision Chains

摘要

Social bot detection is important for maintaining the integrity of online information ecosystems. Although recent graph-based detectors achieve strong predictive performance, their decisions remain difficult to audit because the supporting relational and behavioral signals are encoded in latent representations. Addressing this limitation requires a detector that exposes the evidence behind each prediction and indicates when that prediction may be unreliable. We propose \textbf{VeriBot}, a verification-oriented framework that augments a frozen detector with traceable evidence chains and reliability-aware routing. VeriBot links observable account attributes and relations to decision claims through reproducible transformations, allowing each evidence item to be checked against source data. For routing, VeriBot compares separately trained detector branches and abstains when their predictions conflict. Experiments on Fox8-23 and BotSim-24 show that VeriBot remains competitive with strong baselines while providing grounded evidence for nearly all decisions.