Classification-Error Cost Minimization Strategy: DCMS

Several classification applications such as intrusion detection, biometric recognition, etc. have different costs associated with different classification errors. In such scenarios, the goal is to minimize the cost incurred, and not the classification error rate itself. This paper proposes a Cost Minimization Strategy, dCMS, which when applied to classifiers, provides a boost in the performance by reducing the cost incurred due to classification errors. dCMS is classifier-type independent, however it exploits the statistical properties of the trained classifier. It does not require classifiers to be retrained, which is particularly advantageous in scenarios where the costs vary dynamically. Convincing results are provided which indicate the statistically significant reduction in cost incurred by applying dCMS, in a diverse set of classification scenarios with datasets and classifiers of varying complexities.