An empirical approach toward understanding the effect of misclassification costs on the performance of machine learning algorithms

Evaluating the classifiers and comparing their performance is one of the major concerns of the machine learning community. Empirical measures such as accuracy, entropy and root squared mean error have been used for a long time as the basic criteria in this task. Unfortunately these methods do not take into account the misclassification cost of the class instances. This factor can prove to affect dramatically the performance of the classifier. This paper presents an empirical discussion of the effect of misclassification costs on some machine learning algorithms and investigates the use of ROC analysis to evaluate and compare the performance of these algorithms.