A0268
Title: Subclass discovery from fuzzy decision trees
Authors: Christophe Marsala - Sorbonne Universite (France) [presenting]
Abstract: An approach based on decision trees and fuzzy sets theory is presented to highlight sub-classes in a set of classes. This approach is based on the use of fuzzy sets measures in order to determine non-connected subsets of classes. In a supervised learning approach, the aim is to find convenient features to determine an estimation of the decision frontier that separate different classes. In the case of a (fuzzy) decision tree, this estimation is built by means of a sequence of splits perpendicular to the feature axes. However, it often appears that leaves of the tree labeled with a similar class can be associated with either close regions of examples introducing connected regions of this class in the description space, or distinct and clearly separate regions of examples related to non-connected regions of examples labeled with similar class. This kind of approach is also known as subclass discovery. The aim is to propose an approach that combines a fuzzy decision tree and a clustering approach in order to highlight connected regions. Afterwards, this method is introduced in an explainable artificial intelligence (XAI) approach in order to propose a new way to build surrogates or to find counterfactual examples.
