CrossBot: Drift-Decoupled Adaptation for Cross-Domain Social Bot Detection

Cross-domain social bot detection suffers from substantial performance degradation due to differences in user behaviors, platform characteristics, and bot-generation strategies. Existing unsupervised domain adaptation methods generally model such variations as a single domain discrepancy and apply uniform alignment, which may remove class-discriminative cues or preserve unreliable source decisions. We propose Drift-Decoupled Domain Adaptation (DDDA), a framework that decomposes cross-domain drift into representation-level style drift and decision-level drift. For style drift, DDDA introduces a class-conditional contrastive style alignment mechanism that reduces domain-specific expression differences while preserving discriminative information. For decision drift, DDDA develops a transferability-guided routing mechanism that estimates source decision reliability on unlabeled target domains and adaptively selects appropriate transfer strategies. Extensive experiments on three cross-domain social bot detection datasets demonstrate that DDDA consistently improves target-domain generalization over existing adaptation methods. Further analysis reveals diverse drift patterns across domains, highlighting the importance of drift-aware adaptation.