On Multi-Class Cost-Sensitive Learning

Proceedings of the AAAI Conference on Artificial Intelligence, 21

Volume

Technical Papers

Machine Learning

A popular approach to cost-sensitive learning is to rescale the classes according to their misclassification costs. Although this approach is effective in dealing with binary-class problems, recent studies show that it is often not so helpful when being applied to multi-class problems directly. This paper analyzes that why the traditional rescaling approach is often helpless on multi-class problems, which reveals that before applying rescaling, the {it consistency} of the costs must be examined. Based on the analysis, a new approach is presented, which should be the choice if the user wants to use rescaling for multi-class cost-sensitive learning. Moreover, this paper shows that the proposed approach is helpful when unequal misclassification costs and class imbalance occur simultaneously, and can also be used to tackle pure class-imbalance learning. Thus, the proposed approach provides a unified framework for using rescaling to address multi-class cost-sensitive learning as well as multi-class class-imbalance learning.