Drift-Aware Explainable Social Bot Detection Using Localized Behavioral Drift Monitoring and Adaptive Retraining

This paper presents a drift-aware and explainable framework for social bot detection that addresses the decline in machine learning performance as bot behavior changes over time. The proposed pipeline combines an XGBoost-based baseline classifier, a localized Kolmogorov–Smirnov (KS) drift monitoring mechanism, recency-weighted adaptive retraining, and TreeSHAP-based feature attribution analysis. The framework is evaluated using BotDriftBench v0.8, a synthetic longitudinal benchmark containing 3,000 accounts observed across 20 time windows with 25 features. Experiments use an account-disjoint training, validation, and testing protocol across five random seeds. Results show that the frozen model’s Macro-F1 decreases from 0.9770 before drift to 0.8059 after drift. The proposed localized drift-aware adaptation achieves a temporal Macro-F1 of approximately 0.9167 while requiring fewer retraining events than the oracle adaptation policy. The study also examines changes in feature importance across models to help interpret how adaptation affects the detector’s behavior. All reported experimental results are based on synthetic data. This work contributes an approach for monitoring behavioral drift, adapting detection models, and auditing changes in model feature reliance over time.