Instance-dependent cost-sensitive learning: do we really need it?

Abstract

Traditionally, classification algorithms aim to minimize the number of errors. However, this approach can lead to sub-optimal results for the common case where the actual goal is to minimize the total cost of errors and not their number. To address this issue, a variety of cost-sensitive machine learning techniques has been suggested. Methods have been developed for dealing with both class- and instance-dependent costs. In this article, we ask whether we really need instance-dependent rather than class-dependent cost-sensitive learning? To this end, we compare the effects of training cost-sensitive classifiers with instance- and class-dependent costs in an extensive empirical evaluation using real-world data from a range of application areas. We find that using instance-dependent costs instead of class-dependent costs leads to improved performance for cost-sensitive performance measures, but worse performance for cost-insensitive metrics. These results confirm that instance-dependent methods are useful for many applications where the goal is to minimize costs.