Cost-Sensitive Learning

Abstract

In conventional classification settings, the classifiers generally try to maximize the accuracy or minimize the error rate, both are equivalent to minimizing the number of mistakes in classifying new instances. Such a setting is valid when the costs of different types of mistakes are equal. In real-world applications, however, the costs of different types of mistakes are often unequal. For example, in intrusion detection, the cost of mistakenly classifying an intrusion as a normal access is usually far larger than that of mistakenly classifying a normal access as an intrusion, because the former type of mistakes will result in much more serious losses.