Adaptive Ensemble Learning for Drift Detection and Mitigation in Non-Stationary Data Streams

Non-stationary data streams present a challenge to learning because concept drift decreases the performance of the fixed machine learning models. In this paper, we will suggest a severity-conscious adaptive ensemble framework to support robust online learning with changing data distribution. The approach combines implicit drift detection, adaptive weights of learners and dynamic replacement of learners as a single ensemble model. It is proposed to determine the extent of performance degradation with a new Drift Severity Index (DSI) that can be used to determine the strength of mitigation, and Drift Recovery Time (DRT) is introduced as a new index to measure the speed of adaptation. Moreover, a learner archive system allows managing the recurrent drift effectively by reactivating models. The experimental findings of the synthetic data streams with sudden, gradual and periodic drift prove that the proposed framework maintains stable performance and recovers fast when distributional change is severe. The fact that it has a low per-instance computational complexity renders the approach appropriate in real-time data stream applications.