Possibilistic Feature Selection Method based on Discriminant Power for Class Discrimination

Feature selection is important for the reduction of computing complexity especially in systems dealing with a high amount of data. In this context, various interesting methods of feature selection based on possibilistic model have been proposed. Some of them suffer from system divergence in case of high number of features. In this paper, we propose a feature selection method for class discrimination based on a possibilistic data modeling. In fact, possibilistic modeling is a powerful paradigm, which is able to handle data imperfection. The proposed approach is based on two discriminant levels in order to extract features that are able to discriminate between two classes. The first level depends on the discriminant power of two possibilistic distributions. This level is enhanced by a cross-validation concept. The second level depends on the scoring rate relative of each feature. The approach is validated using synthetic data sets and benchmark data sets. The obtained results show the ability of the proposed approach in selecting the most significant features.