OnlineSIRUOS: An Inverse Random Under and Oversampling, Heterogeneous Ensemble, and Meta-learning Approach for Imbalanced Data Stream Classification

• Vinicios Cainã dos Santos Coelho PUCPR
• Alceu de Souza Britto Jr. PUCPR
• Jean Paul Barddal PUCPR

Resumo

Data streams are potentially unbounded data sequences that are made available rapidly and over time. Due to their pervasiveness, mining data streams has become a major scientific and practical issue. Scenarios involving data streams present multiple challenges, including the requirement for single-pass processing due to constraints on computational resources and the necessity to respond to concept drift over time. Another common trait of several streaming scenarios is class imbalance, that is, a class, often of interest, is majorly outnumbered by others, thus hardening the learning process. This paper introduces Online Stacking Inverse Random Under and Over Sampling (OnlineSIRUOS). This ensemble-based approach combines meta-learning, sampling, and heterogeneous components to address class-imbalanced data stream classification. We evaluated our proposal against existing work tailored for class imbalance in data streams using synthetic and real-world datasets. Experimental results show that our proposal achieves competitive F1 scores in different imbalance ratios and is less computationally intensive than its competitors in processing time and memory consumption. The results also show that our proposal is particularly well-suited for highly imbalanced data streams.