Sélection d’attributs pour arbres de décision flous monotones
In supervised machine learning, it is important to take correctly into account the links between data’s description variables (or attributes) and their class, for instance to build monotonic fuzzy decision trees. In this paper, considering that the attributes and the labeling function are in the form of finite partitions of totally ordered labels, fuzzy versions of the Shannon and Gini rank discrimination measures are introduced, using a fuzzy dominance definition, to allow the selection of attributes monotonically related to the class. Using these measures, a new algorithm for constructing fuzzy decision trees, accounting for a monotonic link between attributes and class, is proposed and experimented on an artificial dataset.