Summarizing Fuzzy Decision Forest by subclass discovery
Résumé
Construction of forest of decision trees method is a popular tool in machine learning because of its good performances in terms of classification power as well as in computational cost. In this paper, we address two problems. The first one concerns the interpretability of a forest. Indeed, comparing to a single decision tree, a forest loose its ability to be easily understandable by an end-user. The second studied problem concerns the size of the forest and hence the memory size and classification time of a forest. We seek for a forest as small as possible that classify nearly as well as a larger forest. In order to solve these two problems, we propose to characterize a forest by discovering different classes of trees regarding their power of classification. These classes are discovered thanks to Forest's algorithm [1] of class segmentation, a variant of the hypersphere classifier [2].
Domaines
Intelligence artificielle [cs.AI]Origine | Fichiers produits par l'(les) auteur(s) |
---|
Loading...