TFIDD: Adaptive dynamics for fast model-agnostic drift detection

Concept drift presents a major challenge for streaming and continual learning systems by degrading predictive performance and destabilizing model adaptation. Existing error-based and distribution-based detectors often struggle to balance sensitivity and stability, leading to excessive false alarms or delayed adaptation in noisy and gradual drift scenarios. We propose TFIDD (Temporal Feature Importance Drift Detector), a novel model-agnostic drift detection framework that monitors temporal feature-importance dynamics rather than relying solely on prediction errors or marginal distribution shifts. The key novelty of TFIDD lies in integrating adaptive Top-k feature selection, temporal smoothing, and confidence-based freezing within a unified feature-driven detection framework, enabling robust drift localization while reducing repeated alarms and preserving interpretability. TFIDD identifies drift through cosine divergence between the current feature representation and a history-weighted baseline using adaptive quantile thresholding. Experiments on synthetic and real-world streams, including SEA, RBF, Electricity, Forest CoverType, Aggarwal, and Gas Sensor datasets, show that TFIDD consistently balances responsiveness and stability across heterogeneous base learners. On the SEA stream, TFIDD achieved a post-drift accuracy of 0.8426 with a false alarm rate of 0.0278, while on the Gas Sensor dataset it maintained a false alarm rate of 0.0023 with stable recovery under annotated drift. Across benchmarks, TFIDD reduced unnecessary alarms compared with highly reactive detectors while preserving competitive detection delay under abrupt, gradual, and recurring drift conditions. By explicitly tracking drift-driving features over time, TFIDD provides both interpretability and computational practicality for long-running adaptive systems.