Using Credal-C4.5 with Binary Relevance for Multi-Label Classification

Abstract

Binary Relevance (BR) is a simple and direct approach to the Multi-Label Classification (MLC). It decomposes the multi-label problem into several binary problems, one per label. It uses an algorithm of traditional supervised classification in order to solve these binary problems. On the other hand, Credal C4.5 (CC4.5) is a modification of the classical C4.5. CC4.5 estimates the probability of the class variable by using imprecise probabilities. In the literature, this new classification algorithm has obtained better results than C4.5 when both have been applied on datasets with class noise. In MLC, since there are not just a class, but multiple labels are disposed, it is more probable that there is intrinsic noise than in traditional classification. From the previous reasons, in this work it is studied the performance of BR using Credal C4.5 as base classifier versus BR with C4.5. It is carried out an experimental study with several muti-label datasets and a considerable number of measures for MLC. This study shows that the performance of BR is improved when it uses CC4.5 as base classifier versus BR with C4.5. In consequence, it is probably suitable to apply imprecise probabilities in Decision Trees within the MLC field too.