Chapter 6
Misclassification costs
6.1 Introduction
In several practical classification tasks, it is not devoid of significance how the performance of a classification model differs for particular classes of the target concept. Models exhibiting seemingly the same performance in terms of overall misclassification rate may vastly differ in actual utility depending on which classes they predict successfully and for which they fail. This is especially true for all kinds of diagnostic or anomaly detection tasks where some model mistakes may be more severe or more tolerable than the others.
To adequately describe the requirements for classification models in such situations, real-valued misclassification costs are used, assigned to particular pairs of predicted and true classes. They can not only provide additional performance criteria for model evaluation, but also—and more importantly—get incorporated into model construction, to make it cost-sensitive.
When discussing misclassification cost incorporation techniques, we will have to refer to classification model quality measures, defined in Chapter 7: the misclassification error, the weighted misclassification error, the mean misclassification cost, and the confusion matrix. This forward reference is unavoidable, because—while it is natural and logical to present model evaluation techniques after cost-sensitive model creation techniques—the former can be only fully justified and understood by referring to the latter.
Get Data Mining Algorithms: Explained Using R now with the O’Reilly learning platform.
O’Reilly members experience books, live events, courses curated by job role, and more from O’Reilly and nearly 200 top publishers.
