SCUT-DS: Learning from Multi-class Imbalanced Canadian Weather Data

Abstract

Learning from multi-class imbalanced data streams with multiple minority classes, and varying degrees of skewed distributions, is an important problem in many real-world applications. However, to date, this aspect has received limited attention in the research community. Rather, the focus is on binary class problems or, alternatively, multi-class scenarios are decomposed into multiple binary sub-problems that are handled separately. Furthermore, the evolving nature of data streams make the task of correctly predicting minority instances challenging. In this paper, we introduce the SCUT-DS approach that combines multi-class synthetic oversampling and cluster-based under-sampling. SCUT-DS is a window-based method that balances the number of incoming instances of all classes directly, as the stream evolves. We present our experimental evaluation against a stream of Canadian weather data, with varying degree of skewed distributions and multiple classes. We demonstrate that our SCUT-DS algorithms consistently improve the recognition rates of the minority instances in this multi-class imbalanced setting. Our results are especially promising for difficult-to-learn minority classes, notably for predicting ice storms and glaze events.