Non-specificity-based Supervised Discretization for Possibilistic Classification

Abstract

AbstractData imperfection and information uncertainty are inevitable in many application domains, including medicine, finance, meteorology, cybersecurity, etc. Sources of uncertainty include the lack of information of the solicited experts, sensor unreliability for automatically collected data, conflicting information sources, etc.. Classification techniques, such as the non-specificity based possibilistic decision trees (NSPDT) have been proposed to deal with such uncertainty. However, most of the introduced possibilistic methods are limited to categorical features, i.e. they can not process data input that is continuous. This paper address this limitation through a supervised possibilistic discretization method, which transforms continuous features into categorical features in datasets where class labels are imperfect and represented by possibility distributions. The proposed method can be used as a pre-processing step to make it possible to use methods such as NSPDT in datasets that initially have continuous features. The proposed non-specificity based discretization algorithm has provided good quality categorical features which have led to obtain NSPDT trees showing an accuracy higher than 80% in most cases. With original datasets, where the instances have certain (i.e. non-possibilistic) class labels, the proposed possibilistic discretization method has shown a robust performance compared to the standard discretization algorithm which is the default approach in this case. The difference in accuracy has not exceeded 5% for all used datasets and all tested classifiers, and in some cases, the possibilistic discretization outperformed the standard discretization.KeywordsDiscretizationPossibility theoryPossibilistic decision treesNon-specificity