Exponentially Weighted Moving Average for Concept Drift Detection With Cautious Learning
Irving D. Estrada-López, Alvaro Eduardo Cordero-Franco · IEEE Access · 2026
Concept drift detection is essential for maintaining the reliability of machine learning systems deployed in dynamic, real-world environments. Drift detectors have become increasingly important for monitoring streaming classification tasks, where models must adapt to continuously evolving data. These detectors must remain both computationally efficient and accurate. This paper introduces the ECDD-CL, Exponentially Weighted Moving Average for Concept Drift Detection with Cautious Learning. The proposed approach extends the original ECDD by incorporating guaranteed in-control performance and a cautious parameter-updating scheme to significantly reduce false alarms—one of the main limitations of conventional detectors. The proposed method was evaluated on synthetic and real datasets and compared with baseline KNN and NB classifiers, the standard ECDD and Drift Detection Method (DDM). Simulation results demonstrate that ECDD-CL achieves competitive accuracy while substantially lowering false alarm rates, providing a stable and computationally efficient framework for reliable drift detection in evolving data streams.