Exploring Label Correlations for Partitioning the Label Space in Multi-label Classification

Recent works on Multi-Label Classification (MLC) present multiple strategies to explore label correlations in a way to improve classifiers performances. However, these works focus only in the traditional local and global approaches, i.e., transforming the original problem into a set of binary local problems, or dealing globally with all classes simultaneously. Very few works have investigated strategies to use label correlations in order to partition the label space in a different ways. While in local partitions several binary classifiers are used (one per label), global partitions use only one classifier to deal with all labels. On the contrary, here we propose a strategy that explores the correlations between labels to partition the label space aiming to find partitions in-between (hybrid) the local and global ones. We believe in-between local and global partitions better cluster similar labels, improving the multi-label classifiers ability to explore label correlations. We compared the hybrid partitions with global, local and random generated partitions. Our experimental results showed that the hybrid partitions lead to competitive results and, in general, were slightly better than global and local partitions. The random partitions were also competitive with the global and local partitions, showing that the current local and global approaches still need improvements in order to consider label correlations.