Label Clustering for Particle Swarm Optimisation based Multi-Label Feature Selection

Feature selection is an important pre-processing step for multi-label classification, which selects a small subset of complementary features and improves the classification performance. Particle Swarm Optimisation (PSO) has recently been widely applied to achieve multi-label feature selection. However, when the number of labels is large, existing PSO-based multilabel approaches are computationally intensive. In multi-label classification, many labels usually appear together (for example, "sea" and "ship" labels), which are regarded as similar labels. This work proposes to group similar labels together to form a label cluster. By considering each label cluster as a super label, the number of labels is significantly reduced, and thus using the super labels can significantly reduce the computational cost. The experimental results on nine real-world datasets show that the proposed label clustering mechanism can maintain or even improve the multi-label classification performance while being much more efficient and selecting smaller numbers of features than using a large number of original labels.