Uncertainty-Driven Online Deep Ensemble for Imbalanced Drifting Data Stream Regression

Concept drift presents significant challenge in many real-world data stream regression tasks, particularly in industrial applications. This challenge becomes even more complex when concept drift is accompanied by data imbalance. To address this issue, we propose an uncertainty-driven online deep ensemble framework (UODEF) which leverages deep ensemble embedding with dynamic weight adjustment based on both performance and time factor. Specifically, UODEF incorporates epistemic uncertainty-driven drift detection to detect concept drift, and retrains new base learners utilizing historical information to effectively handle recurring drifts. To address data imbalance, an uncertainty-driven adaptive oversampling technique is developed to dynamically identify and oversample rare samples, ensuring that new base learners are unbiased. Experiments on a real-world dataset from a diesel hydrofining process, collected in a petrochemical workshop, demonstrate that UODEF outperforms three recently proposed methods for imbalanced drifting data stream regression.