Class Segmentation to Improve Fuzzy Prototype Construction: Visualization and Characterization of Non Homogeneous Classes

Jason Forest (1) , Maria Rifqi (1) , Bernadette Bouchon-Meunier (1)

1 MALIRE - Machine Learning and Information Retrieval

In this paper, we present a new method to construct fuzzy prototypes of heterogeneous classes, in a supervised learning context. Heterogeneous classes are classes where the coexistence of far behaviours can be observed. Our approach consists in two stages. The first one enables to discover, in an original method, the different behaviours within a class by decomposing it in subclasses. In the second stage, we construct a fuzzy prototype for each subclass by using typicality degrees. Thanks to this decomposition of a class and to this characterization of typical behaviours, we propose an intuitive summarization of a class. We illustrate the advantages of our method on both artificial and real dataset.

Domaines: Intelligence artificielle [cs.AI]

Jason Forest, Maria Rifqi, Bernadette Bouchon-Meunier. Class Segmentation to Improve Fuzzy Prototype Construction: Visualization and Characterization of Non Homogeneous Classes. 2006 IEEE International Conference on Fuzzy Systems, Jul 2006, Vancouver, Canada. pp.555--559. ⟨hal-01072669⟩