Misclassification costs
Paweł Cichosz · 2015
This chapter focuses on the basic instance-independent cost matrix and cost vector representations, but possible extensions of the presented techniques to instance-specific costs are suggested where appropriate. It reviews several general techniques that can be used to make classification algorithms cost-sensitive. One approach to incorporating misclassification costs that is not algorithm specific, but applicable to a wider class of classification algorithms, is based on instance weighting. Instance weighting or instance resampling can deal with simplified cost matrices only. The instance relabeling technique turns out to work better than the minimum-cost rule for decision trees on the original four-class dataset only if class probabilities are estimated with bagging. The experimental procedure for drawing reliable conclusions about the capabilities of the techniques of misclassification cost incorporation should include multiple algorithms, multiple datasets, random costs, and reliable evaluation.