Many research work in cost-sensitive learning focused on binary class problems and assumed that the costs are precise. But real-world applications often have multiple classes and the costs cannot be obtained precisely. It is important to address these issues for cost-sensitive learning to be more useful for real-world applications. This paper gives a short introduction to cost-sensitive learning and then summaries some of our previous work related to the above two issues: (1) The analysis of why traditional Rescaling method fails to solve multi-class problems and our method Rescale
. (2) The problem of learning with cost intervals and our CISVM method. (3) The problem of learning with cost distributions and our CODIS method.