Adaptive Ensemble Undersampling-Boost: A novel learning framework for imbalanced data

As one of the most challenging and attractive problems in the pattern recognition and machine intelligence field, imbalanced classification has received a large amount of research attention for many years. In binary classification tasks, one class usually tends to be underrepresented when it consists of far fewer patterns than the other class, which results in undesirable classification results, especially for the minority class. Several techniques, including resampling, boosting and cost-sensitive methods have been proposed to alleviate this problem. Recently, some ensemble methods that focus on combining individual techniques to obtain better performance have been observed to present better classification performance on the minority class. In this paper, we propose a novel ensemble framework called Adaptive Ensemble Undersampling-Boost for imbalanced learning. Our proposal combines the Ensemble of Undersampling (EUS) technique, Real Adaboost, cost-sensitive weight modification, and adaptive boundary decision strategy to build a hybrid algorithm. The superiority of our method over other state-of-the-art ensemble methods is demonstrated by experiments on 18 real world data sets with various data distributions and different imbalance ratios. Given the experimental results and further analysis, our proposal is proven to be a promising alternative that can be applied to various imbalanced classification domains.