Feature Selection for Hierarchical Multi-label Classification

Abstract

In this work we study how conventional feature selection methods can be applied to Hierarchical Multi-label Classification Problems. In Hierarchical Multi-label Classification, instances can belong to two or more classes (labels) simultaneously, where such classes are hierarchically structured. Feature selection plays an important role in Machine Learning classification tasks, once it can effectively reduce the dataset dimensionality by removing irrelevant and/or redundant features, improving classification accuracy. Although many relevant real-world problems are from the hierarchical and multi-label domains, the majority of the related researches address the feature selection task focusing on single-label problems. In many works, even when the proposal deals with multi-label problems, the classes are not associated with a hierarchical structure. Therefore, in this work we study how feature selection can be applied in the Hierarchical Multi-label Classification context. For this, we propose four hierarchical strategies combining the Binary Relevance (BR) and Label Powerset (LP) multi-label transformations with the attribute evaluators ReliefF (RF) and Information Gain (IG). We tested our strategies on 10 real-world datasets from the functional genomic field, commonly used in Hierarchical Multi-label Classification works. As main results, three of the four proposed strategies produced some relevant subsets of features, while keeping predictive performances in comparison to the use of the complete set of features.