On Unifying Probabilistic Fuzzy and Possibilistic Rejection Based Classifiers英文文献.pdfVIP

On Unifying Probabilistic / Fuzzy and Possibilistic Rejection-Based Classifiers Carl Fr6ticot Universit6 de La Rochelle, Laboratoire dInformatique et dImagerie Industrielle, 17042 La Rochelle Cedex 1, FRANCE Phone: +33 546 458 234, Fax: +33 546 458 242, e-mail: cfrelico@gi.univ-lr.fr Abstract. This paper aims at unifying the presentation of two-fold rejection- based pattern classifiers. We propose to define such a classifier as a couple of labelling and hardening functions which are independent in some way. Within this framework, crisp and probabiIistic / fuzzy rejection-based classifiers are shown to be particular cases of possibilistic ones. Classifiers with no reject option remains particular cases of rejection-based ones. Examples of so-defined classifiers are presented and their ability to deal with the reject problem is shown on artificial and real data sets. 1 Pattern Classification The pattern classification problem can be defined as follows. Let x = (xl, x2, ..., xp) ~ be a pattern described by p features, and ~ = {a~l,a32, .-.,~c} be a set of c classes. A classifier performs a mapping: NP -+ L.c a class labels space, x ~ t(x) = (ll(x),12(x), ...,lc(X)). Three sets of class label vectors in Nc may be defined as in [1]: I. Lpc = [0, 1] ~, i.e. the unit hypercube in ~c ; l(x) E Lpc is a possibilistic label vector, e.g. l(x) = (0, 0.7, 0.5) t. Some authors exclude the origin of the unit hypercube from Lpc, e.g. in [