International Journal of Computer Theory and Engineering

Abstract

In such real data mining applications as medical diagnosis, fraud detection and fault classification, and so on, the two problems that the error cost is expensive and the reject cost is class-dependent are often encountered. In order to overcome those problems, firstly, the general mathematical description of the Binary Classification Problem with Error Cost and Class-dependent Reject Cost (BCP-EC2RC) is proposed. Secondly, as one of implementation methods of BCP-EC2RC, the new algorithm, named as Cost-sensitive Support Vector Machines with the Error Cost and the Class-dependent Reject Cost (CSVM-EC2RC), is presented. The CSVM-EC2RC algorithm involves two stages: estimating the classification reliability based on trained SVM classifier, and determining the optimal reject rate of positive class and negative class by minimizing the average cost based on the given error cost and class-dependent reject cost. The experiment studies based on a benchmark data set illustrate that the proposed algorithm is effective.