Résumé: In the age of big data, the number of class labels is increasing rapidly and there exists a hierarchical structure between different class labels. In the present paper, we revisit the existing granular computing approach to hierarchical classification. By revealing some limitations of approximation capacity, we develop a novel model for hierarchical classification. Then, we present a formal approach to feature selection for hierarchical decision tables by using fuzzy rough set theory. Correspondingly, an algorithm using relative discernibility relation is designed to select relevant feature subsets. Considering the fact that real data may vary dynamically with time, we also propose an incremental approach to hierarchical feature selection by using fuzzy rough set technique. An incremental algorithm for hierarchical feature selection is provided based on the sibling strategy. The experimental results demonstrate that the proposed approach is feasible and valid.

Langue: Anglais