Dynamic Windowing Strategies for Concept Drift Type Classification in Data Streams

Detecting and classifying concept drift in data streams is essential for maintaining the accuracy and reliability of machine learning models deployed in dynamic environments. Traditional drift detection methods often apply a single windowing strategy across all drift types, limiting their ability to distinguish between sudden, gradual, incremental drifts. In this paper, we propose a novel approach that leverages tailored windowing strategies to enhance the performance of concept drift classifiers. Specifically, we apply sliding windows for gradual drift, tumbling windows for sudden drift, expanding windows for incremental drift, aligning each strategy with the characteristics of the respective drift type. By generating synthetic data streams using these customized windowing strategies, we train a drift classifier capable of both detecting and accurately classifying different types of drift in real-time data streams. Experimental results show that our approach significantly improves drift detection accuracy and classification performance compared to conventional windowing methods. These findings highlight the potential of tailored pre-training strategies in adaptive systems and offer a pathway for developing more robust and context-aware drift detection solutions.