Adaptive Concept Drift Detection Using Bayesian Neural Networks and Wasserstein Distance

Concept drift in streaming data poses a significant challenge to the stability and performance of deep learning models. This study explores an uncertainty-based drift detection approach leveraging Wasserstein Distance in a Bayesian Neural Network (BNN) framework. By continuously monitoring uncertainty variations across batches, the proposed method effectively identifies changes in data distribution. Experimental results demonstrate that lower drift thresholds lead to increased drift detection sensitivity, while higher thresholds result in more conservative drift identification. The findings validate Wasserstein Distance as a robust metric for real-time concept drift detection, providing a balance between model adaptability and stability. The study further highlights the necessity of adaptive learning strategies to handle evolving data streams.