A new classifier for categorical data based on a possibilistic estimation and a novel generalized minimum-based algorithm

Abstract

In this paper, we propose NPC _c , a new Naïve Possibilistic Classifier for categorical data. The proposed classifier relies on the Bayesian structure of the Naïve Bayes Classifier for categorical data (NBC _c ) which stands for an interesting pattern when dealing with discrete attributes. However, unlike NBC _c , the proposed NPC _c is based on the possibilistic formalism as an efficient fuzzy-sets-based alternative to the probabilistic one when handling uncertain data. Distinctively, we use the possibilistic approach to estimate beliefs from categorical data and a Generalized Minimum-based classification algorithm (G-Min) as a novel algorithm to make decision from possibilistic beliefs. Experimental evaluations on 12 datasets taken from University of California Irvine (UCI) and containing all categorical data, confirm the effectiveness of the proposed new G-Min-based NPC _c . With the used datasets, the proposed classifier outperforms the commonly-used classifiers for categorical data including NBC _c , C4.5-based decision tree and RIPPER-based classifier. Moreover, it outperforms the two versions of NPC _c using commonly-used possibilistic classification algorithms which are based on respectively, the product and the minimum operators. Consequently, we prove the efficiency of the possibilistic approach together with the G-Min algorithm for the classification of categorical data.