Adaptive Classifier to Address Concept Drift in Imbalanced Data Streams

Concept drift adaptation is essential for maintaining the accuracy of machine learning models when facing changing data distributions in streaming environments. Class imbalance, a common issue in real-world datasets, further compounds the challenge of concept drift detection and adaptation. This paper proposes a comprehensive approach that focuses on effectively identifying and handling both concept drift and class imbalance to ensure the reliability and performance of predictive models. The approach involves two stages. First, a concept drift detection mechanism is employed to monitor the changing data distribution among the historical context of class distributions and adapts its sensitivity to class-wise changes. This minimizes the risk of false alarms while capturing subtle shifts in minority class distributions. Second, for concept drift adaptation, it uses ensemble-based resampling technique that dynamically adjusts the balance of classes by oversampling the minority class instances in a controlled manner. The approach is evaluated on synthetic and real-world streaming datasets with varying degrees of class imbalance and concept drift. Experimental results determine the advantage of the proposed method in comparison to baseline techniques in terms of detection accuracy and adaptability to changing conditions, particularly for underrepresented classes.