Unsupervised Concept Drift Detection and Adaptation Based on Random Forest

Concept drift, a prevalent challenge in dynamic data environments, occurs when the statistical properties of data distributions change over time, leading to a degradation in classifier performance. Traditional concept drift adaptation techniques often rely on labeled data for model retraining, which may not be practical in real-world streaming scenarios due to the high cost and limited availability of labeled samples. In this paper, we propose an unsupervised concept drift detection and adaptation framework based on random forest (RF) and Kullback-Leibler (KL) divergence, referred to as RFKL. It reduces reliance on labeled data and enhancing classifier robustness in dynamic settings.