… Séminaires de l’UMR ITAP …
Lundi 23 septembre 2013 à 10 heures, Irstea Montpellier, bât. Minéa , salle Orient
Using fuzzy measures to characterize subset of attributes
par Javier Murillo, CIFASIS (Centro Internacional Franco Argentino de Ciencias de la Información y de Sistemas – Rosario, Argentina)
Classification is required in many human activities. Supervised learning consists in designing a classifier from a set of training samples for which the description, a set of attributes, but also the category membership are known. Attribute selection is a key step in the process of classifier design.
Most of attribute selection techniques rely on individual attribute evaluation, assuming independence. New proposals aim at selecting subsets of variables according to their collective behavior. Fuzzy measures are useful to weigh partial contributions, to be integrated using a Choquet integral. Shapley index, and its generalization, is used to summarize the contribution of each subset to the classification process.
The whole process is applied to benchmark data and the results are compared to other approaches.