From Bayesian classifiers to possibilistic classifiers for numerical data - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2010

From Bayesian classifiers to possibilistic classifiers for numerical data

Résumé

Naïve Bayesian classifiers are well-known for their simplicity and efficiency. They rely on independence hypotheses, together with a normality assumption, which may be too demanding, when dealing with numerical data. Possibility distributions are more compatible with the representation of poor data. This paper investigates two kinds of possibilistic elicitation methods that will be embedded into possibilistic naïve classifiers. The first one is derived from a probability-possibility transformation of Gaussian distributions (or mixtures of them), which introduces some further tolerance. The second kind is based on a direct interpretation of data in fuzzy histogram or possibilistic formats that exploit an idea of proximity between attribute values in different ways. Besides, possibilistic classifiers may be allowed to leave the classification open between several classes in case of insufficient information for choosing one (which may be of interest when the number of classes is large). The experiments reported show the interest of possibilistic classifiers.

Dates et versions

hal-03349427 , version 1 (20-09-2021)

Identifiants

Citer

Myriam Bounhas, Khaled Mellouli, Henri Prade, Mathieu Serrurier. From Bayesian classifiers to possibilistic classifiers for numerical data. 4th International Conference on Scalable Uncertainty Management (SUM 2010), Sep 2010, Toulouse, France. pp.112-125, ⟨10.1007/978-3-642-15951-0_15⟩. ⟨hal-03349427⟩
8 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More