Improving classifier utility by altering the misclassification cost ratio

This paper examines whether classifier utility can be improved by altering the misclassification cost ratio (the ratio of false positive misclassification costs to false negative misclassification costs) associated with two-class datasets. This is evaluated by varying the cost ratio passed into two cost-sensitive learners and then evaluating the results using the actual (or presumed actual) cost information. Our results indicate that a cost ratio other than the true ratio often maximizes classifier utility. Furthermore, by using a hold out set to identify the "best" cost ratio for learning, we are able to take advantage of this behavior and generate classifiers that outperform the accepted strategy of always using the actual cost information during the learning phase.