Feature selection refers to data reduction process by selecting the minimal subsets of features which are effective to preserve the meaning of the features and rarely dependent on other features. Fuzzy-rough set-based feature selection is a beneficial technique which not only satisfies these conditions but also can deal with imprecision and uncertainty. Many methods have been proposed for feature selection problem; however, most of them are able to find only one minimal data reduction while a dataset can have several minimal reducts. In this paper, we propose a Fuzzy-rough set-based feature selection, using particle swarm optimization (PSO) technique, able to find various minimal data reductions. The main contribution of this paper includes using a ring topology for a binary version of the PSO, utilizing the fuzzy-rough dependency degree as fitness. In addition, we present a new velocity updating rule. In order to obtain the efficiency of the proposed method, we compare it with some other meta-heuristic methods using 10 well-known UCI data sets. The results show that the performance of the fuzzy rough-based feature selection can be improved using this method for finding various data reductions.

Finding out information about features is one of the main goals of feature selection. In that case there is not particularly care about the resulting classification accuracy, but we are interested in maximizing dependency degree. A data set can have several minimal data reductions; however, most of the methods are able to find only one minimal data reduction which is not beneficial. In this paper, we propose a feature selection method based on modified Ant Colony Optimization algorithm (ACO). The main contribution of this paper includes using fuzzy-rough gain ratio as heuristic information and new rules for pheromone updating in ACO. Unlike most of the methods which find only one minimal reduction, this method is able to find various minimal data reductions. The proposed method is compared with three other meta-heuristic methods. The results show that our approach is very successful in finding various minimal data reductions.

No abstract