Communication Dans Un Congrès Année : 2024

Harnessing Superclasses for Learning from Hierarchical Databases

Résumé

In many large-scale classification problems, classes are orga-nized in a known hierarchy, typically represented as a tree expressingthe inclusion of classes in superclasses. We introduce a loss for this typeof supervised hierarchical classification. It utilizes the knowledge of thehierarchy to assign each example not only to a class but also to all encom-passing superclasses. Applicable to any feedforward architecture with asoftmax output layer, this loss is a proper scoring rule, in that its expecta-tion is minimized by the true posterior class probabilities. This propertyallows us to simultaneously pursue consistent classification objectives be-tween superclasses and fine-grained classes, and eliminates the need fora performance trade-off between different granularities. We conduct anexperimental study on three reference benchmarks, in which we vary thesize of the training sets to cover a diverse set of learning scenarios. Ourapproach does not entail any significant additional computational costcompared with the loss of cross-entropy. It improves accuracy and re-duces the number of coarse errors, with predicted labels that are distantfrom ground-truth labels in the tree.
Fichier principal
Vignette du fichier
main.pdf (837.64 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04695348 , version 1 (22-11-2024)

Identifiants

Citer

Nicolas Urbani, Sylvain Rousseau, Yves Grandvalet, Leonardo Tanzi. Harnessing Superclasses for Learning from Hierarchical Databases. European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD 2024), Sep 2024, Vilnius (Lituanie), Lithuania. pp.247-265, ⟨10.1007/978-3-031-70359-1_15⟩. ⟨hal-04695348⟩
18 Consultations
6 Téléchargements

Altmetric

Partager

More