G-mean Weighted Classification Method for Imbalanced Data Stream with Concept Drift

Concept drift and class imbalance in data stream seriously degrade the performance and stability of the traditional data stream classification algorithms. To solve this issue in binary classification of data stream, an online G-mean weighted ensemble classification method for imbalanced data stream with concept drift termed OGUEIL is proposed. It exploits the online update mechanism of component classifiers’ weights to modify block-based ensemble algorithms, combining the hybrid resampling and adaptive sliding window algorithm. OGUEIL is based on the ensemble learning framework that once a new instance reaches, each component classifier in the ensemble and its weight are correspondingly updated online, and the minority class instance is randomly oversampled at the same time. Particularly, each component classifier determines its weight according to the G-mean performance on several recently incoming instances, where G-mean of each component classifier is calculated based on the time decay factor increment. At the same time, OGUEIL periodically constructs a balanced dataset according to the data in the current sliding window and trains a new candidate classifier, then adds it to the ensemble based on specific conditions. The experimental results on both real-world and synthesized datasets show that the comprehensive performance of the proposed method outperforms other baseline algorithms.