Feature Ranking with Relief for Multi-label Classification: Does Distance Matter?

Abstract

In this work, we address the task of feature ranking for multi-label classification (MLC). The task of MLC is to predict which labels from a maximal predefined label set are relevant for a given example. We focus on the Relief family of feature ranking algorithms and empirically show that the definition of the distances in the target space used within Relief should depend on the evaluation measure used to assess the performance of MLC algorithms. By considering different such measures, we improve over the currently available MLC Relief algorithm. We extensively evaluate the resulting MLC ranking approaches on 24 benchmark MLC datasets, using different evaluation measures of MLC performance. The results additionally identify the mechanisms of influence of the parameters of Relief on the quality of the rankings.