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

Gatto, Elaine Cecilia; Ferrandin, Mauri; Cerri, Ricardo

doi:10.1109/ijcnn52387.2021.9533331

Cited by 5 publications

(4 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Hybrid Partitions for Multi-Label Classification with Jaccard (HPML-J) is an extension of the method proposed by Gatto et al [51]. Here we extended the experiments and analyses, comparing the generated hybrid partitions with the ones generated by global, local, random, and oracle-based methods and also including more datasets.…”

Section: Hpml-jmentioning

confidence: 98%

“…In this paper, we throughout extended the previous experiments of Gatto et al [51] and also implemented three other different HPML-J variants. The H JM a version chooses the optimal partition based on the Macro-F1 metric of a classifier.…”

Section: Hpml-jmentioning

confidence: 99%

“…To choose the best partition, Gatto et al [51] used the Macro-F1 measure. As a result, the strategy selects the partition with the greatest improvement in the Macro-F1 value of the classifier.…”

Section: Hpml-jmentioning

confidence: 99%

See 2 more Smart Citations

Multi-Label Classification with Label Clusters

Gatto

Ferrandin

Cerri

2023

Preprint

Self Cite

View full text Add to dashboard Cite

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.

show abstract

Section: Hpml-jmentioning

confidence: 98%

Section: Hpml-jmentioning

confidence: 99%

See 1 more Smart Citation

Multi-Label Classification with Label Clusters

Gatto

Ferrandin

Cerri

2023

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…One such strategy is proposed in Gatto et al [28], where Jaccard indexes are used to map the label space into hybridized partitions. This helps to combine the benefits of binary and label power set methods by allowing label dissimilarity to drive the decision on assuming label dependencies.…”

Section: Ensembles For Multi-label Datamentioning

confidence: 99%