BotRoute: Dual-Branch Selective Routing for Social Bot Detection

摘要

Social bot detection is essential for preserving the reliability and integrity of online social platforms. As bot evasion strategies evolve, advanced bot accounts increasingly resemble genuine users in behavioral, textual, and relational patterns, making them difficult to classify reliably with a single detector. However, existing detectors typically apply a uniform decision rule to all nodes, overlooking a small subset of hard cases in which camouflage, noisy links, and heterophilous interactions distort different evidence sources and lead to conflicts among detection views. We observe that such disagreement cases reveal a form of expert complementarity: the strongest detector overall is not necessarily the most reliable for every node where detection views disagree, and experts with different inductive biases can correct different subsets of errors. Motivated by this observation, we propose BotRoute, a dual-branch selective expert handoff framework for social bot detection. BotRoute maintains a graph-based detector as the default expert, activates auxiliary experts only on disagreement nodes, and evaluates each candidate handoff using both probability-level utility and structure-based reliability. Extensive experiments demonstrate that BotRoute improves social bot detection performance and effectively converts expert complementarity into more reliable node-level decisions.