Abstract
Feature drift is a subtype of data distribution drift that occurs when the statistical significance of input features changes over time, despite the overall decision boundary remaining stable. This phenomenon can cause a subtle degradation in model accuracy in streaming environments. In this paper, we propose a new model-aware method called feature importance-driven drift detection (FIDD). Rather than relying on classification error signals, FIDD tracks changes in feature importance rankings obtained from LASSO regression across neighbouring data fragments. As it observes the dynamics of feature importance instead of global label shifts, this method is particularly suited to detecting subtle shifts in data distribution. Experimental evaluation on both synthetic and real-world data (including different types of drift, such as abrupt, gradual, incremental and recurrent) shows that FIDD achieves higher accuracy consistently and produces significantly fewer false alarms than standard drift detectors (e.g., DDM, EDDM and ADWIN). Furthermore, FIDD is robust to labelling noise and computationally efficient, which makes it a practical and interpretable solution for adaptive learning in real-time applications.