Parameter Distribution Ensemble Learning for Sudden Concept Drift Detection.

Concept drift is a big challenge in data stream mining (including process mining) since it seriously decreases the accuracy of a model in online learning problems. Model adaptation to changes in data distribution before making new predictions is very necessary. This paper proposes a novel ensemble method called E-ERICS, which combines multiple Bayesian-optimized ERICS models into one model and uses a voting mechanism to determine whether each instance of a data stream is a concept drift point or not. The experimental results on the synthetic and classic real-world streaming datasets showed that the proposed method is much more precise and more sensitive (shown in F1-score, precision, and recall metrics) than the original ERICS models in detecting concept drift, especially a sudden drift.