No abstract

Multi-Label Classification is the task of simultaneously predicting a set of labels for an instance. Typically, two approaches are used: global, which trains a single classifier to deal with all classes at once, and local, which divides the problem into many binary problems. In both approaches, learning label correlations is still an open issue. In this paper, we propose a method to cluster the label space and find partitions of disjoint correlated labels called hybrid partition, which can be considered in-between the local and global, combining the benefits of both. Our proposal entails: i) clustering the label space based on label correlations; ii) generating and validating the resulting hybrid partitions; iii) selecting the optimal partitions between them; and iv) evaluating their performances. Since we are also interested in analyzing how hybrid partitions are close to the best possible partitions, we compared our results with the ones obtained using partitions found by an oracle, an exhaustive, and a random search. The oracle and exhaustive strategies generate all possible label partitions for a dataset. While the oracle chooses the best one based on the test set alone, the exhaustive choices are based on validation. Experiments on eight well-known multi-label datasets revealed that both our proposal and the random partitions achieve superior or competitive results compared with the others. This indicates that the conventional global and local approaches may not accurately account for label correlations, requiring significant improvements. In this case, clustering the label space can be a solution.

No abstract

No abstract