On the evaluation of example-dependent cost-sensitive models for tax debts classification

Abstract

Example-dependent cost-sensitive classification methods are suitable to many real-world classification problems, where the costs, due to misclassification, vary among every example of a dataset. Tax administration applications are included in this segment of problems, since they deal with different values involved in the tax payments. To help matters, this work presents an experimental evaluation which aims to verify whether cost-sensitive learning algorithms are more cost-effective on average than traditional ones. This task is accomplished in a tax administration application domain, what implies the need of a cost-matrix regarding debt values. The obtained results show that cost-sensitive methods avoid situations like erroneously granting a request with a debt involving millions of reals. Considering the savings score, the cost-sensitive classification methods achieved higher results than their traditional method versions.