R3D: an approach combining cost-sensitive and insensitive classifiers to achieve more balanced scores

Abstract Cost-sensitive classifiers have demonstrated to be more cost-effective than traditional classifiers in many real-world classification problems by avoiding high amounts of money loss. This is usually measured by a savings score. On the other hand, these classifiers present, in some cases, a reduction on conventional performance scores such as accuracy, f-score, precision and recall rates. Government applications are included in this segment of problems but, on public domains, the pursuit to the lowest cost is not an exclusive priority. In this light, this work proposes an approach which uses traditional and example-dependent cost-sensitive supervised learning in order to classify requests for revision of a Brazilian tax administration service. This hybrid approach has achieved a better trade-off when considering cost savings, accuracy, f-score, precision and recall in accordance to the business and data rules.