Structuring the Output Space in Multi-label Classification by Using Feature Ranking

Abstract

Motivated by the increasing interest for the task of multi-label classification (MLC) in recent years, in this study we investigate a new approach for decomposition of the output space with the goal to improve the predictive performance. Namely, the structuring of the output/label space is performed by constructing a label hierarchy and then approaching the MLC task as a task of hierarchical multi-label classification (HMLC). Our approach is as follows. We first perform feature ranking for each of the labels separately and then represent each of the labels with its corresponding feature ranking. The construction of the hierarchy is performed by the (hierarchical) clustering of the feature rankings. To this end, we employ four clustering methods: agglomerative clustering with single linkage, agglomerative clustering with complete linkage, balanced k-means and predictive clustering trees. We then use predictive clustering trees to estimate the influence of the constructed hierarchies, i.e., we compare the predictive performance of models without exploiting the hierarchy and models using hierarchies constructed using label co-occurrences or per label feature rankings. Moreover, we investigate the influence of the hierarchy in the context of single models and ensembles of models. We evaluate the proposed approach across 8 datasets. The results show that the proposed method can yield predictive performance boost across several evaluation measures.