Problems of classification in a fuzzy environment (Q789127): Difference between revisions

The aim of the present paper is to set a classification model in a fuzzy environment instead of a random environment. In this way, the paper starts assuming that the considered ''classes'' are not precise but rather fuzzy sets, so that the belonging of each of the items to a concrete ''class'' is known only as a possibility distribution (in Zadeh's sense). As well as the discriminant analysis usually considers the decision rules which are optimal according to the Bayes principle, the rules to be considered in the fuzzy model would be ''optimal'' according to some principle. This principle is to be based on the concept of Bayes profit, which is introduced herein in terms of the Sugeno's integral with respect to a fuzzy measure. Then, from such a concept, authors state a principle of choice among membership functions, which involves certain similarities with the Bayes principle in a random environment. The final purpose of this paper is to find ''optimal'' rules in the sense above. This purpose is achieved as follows: First, it is proved that the only assumption of the existence of a possibility distribution on the product of the item and class spaces allows for an induction of fuzzy measures on each of them, as well as the corresponding conditional fuzzy measures; an optimal membership function is defined for each ''class'' in terms of the conditional fuzzy measures; finally, a short study for the ''admissibility'' of the definition above is given. The paper concludes suggesting two generalizations for the fuzzy model: a first one, consisting in incorporating probabilistic elements, and a second one, concerning the introduction of a cost structure.