Comparative study of cost-sensitive classifiers

The authors briefly review the theory of cost-sensitive learning, and the existing cost-sensitive learning algorithms. The authors categorize cost-sensitive learning algorithms into direct cost-sensitive learning and cost-sensitive meta-learning, which converts cost-insensitive classifiers into cost-sensitive ones. The authors also propose a simple yet general and effective meta-learning method called Empirical Threshold Adjusting (ETA for short). The authors evaluate the performance of various cost-sensitive meta-learning algorithms including ETA. ETA almost always produces the lowest misclassification cost, and is least sensitive to the misclassification cost ratio. Other useful conclusions on cost-sensitive meta-learning methods are drawn.