A diversity-based method for class-imbalanced cost-sensitive learning

It is often the case that datasets are imbalanced in the real world. In this situation, it is minimizing misclassification costs rather than classification accuracy that is the primary goal of classification algorithms. To tackle this problem and improve the performance of classifiers, sampling is widely employed. In this paper, we propose a new diversity-based under-sampling technique for class-imbalanced datasets. The key idea is to balance a data set by choosing only the potential informative samples of the majority class according to diversity of class probability calculation. The experimental results on 5 class-imbalanced datasets show that our method performs better than two existing sampling techniques in terms of total misclassification costs.