Hierarchical Cost-Sensitive Techniques for Class Imbalance Learning
Huan Xu · 2021
In the traditional classification, the classification of a new example is based on maximum of posterior probability (minimizing the error rate). This implicitly assumes that all misclassification errors cost equally. However, in real-world applications, such as face recognition, medical diagnosis, spam filtering, these assumptions are not reasonable, since costs between different misclassification errors can be quite different. Cost-sensitive learning has been developed for many years to solve this problem, but most of existing methods are developed for two-class cases, yet how to extend these methods and theories to multi-class cases is not straightforward. While several multiclass cost-sensitive algorithms have been proposed, they cannot leverage the minority classes efficiently since they treat each minority class individually. Aiming at addressing this issue, we first investigate multi-class cost-sensitive kernel logistic regression (mcKLR), which is derived from Bayes decision theory and cost-sensitive learning theory. Furthermore, based on the ideas of mcKLR, we also propose a hierarchical approach to cost-sensitive, and study its effectiveness in the context of kernel logistic regression, the experimental results justify the effectiveness and efficiency of our algorithm.