Learning with cost intervals

Existing cost-sensitive learning methods work with unequal misclassification cost that is given by domain knowledge and appears as precise values. In many real-world applications, however, it is difficult to have a precise cost information since the user maybe only knows that one type of mistake is much more severe than another type, yet not possible to give a precise description. We claim that, in such situations, it is more meaningful to work with cost intervals instead of a precise cost value. We propose the CISVM method, a support vector machine that can work with cost interval information. Experimental results show that when there is only cost interval information available, CISVM is superior to training a standard cost-sensitive SVM by using minimal cost, mean cost and maximal cost. Key words: Cost-sensitive learning, cost interval, unequal cost

• Type: Conference
• Year: 2010
• Where: KDD
• Authors: Xu-Ying Liu, Zhi-Hua Zhou
• Keywords: Cost Interval, Cost Interval Information, Data Mining, KDD 2010, Precise Cost