Multi-label classification with label clusters

Multi-label classification is the task of simultaneously predicting a set of labels for an instance, with global and local being the two predominant approaches. The global approach trains a single classifier to handle all classes simultaneously, while the local approach breaks down the problem into multiple binary problems. Despite extensive research, effectively capturing label correlations remains a challenge in both methods. In this paper, we introduce an approach that clusters the label space to create hybrid partitions (disjoint correlated label clusters), striking a balance between global and local strategies while leveraging both advantages. Our approach consists of (i) clustering the label space based on correlations, (ii) generating and validating the resulting hybrid partitions, (iii) selecting the best partitions, and (iv) evaluating their performance. We also compare our approach against an oracle, exhaustive search, and random search to assess how closely our hybrid partitions approximate the best possible partitions. The oracle selects the best partition using the test set, while the exhaustive approach relies on validation data. Experiments conducted on multiple multi-label datasets demonstrate that our method, along with random partitions, achieves results that are superior or competitive compared to traditional global and local approaches, as well as the state-of-the-art Ensemble of Classifier Chains. These findings suggest that conventional methods may not fully capture label correlations, and clustering the label space offers a promising solution.