Many real datasets contains attributes that are not relevant to the task at hand, even more, they are not correlated with the output class. Some of the attributes may be redundant, and others may be irrelevant and noisy. The process of features selection aims to find those attributes. In order to improve the process of feature selection, in this paper we introduce a novel metric for fuzzy-rough feature selection (FRFS). The new fuzzy similarity metric is evaluated using Lukasiewicz and Kleene-Dienes fuzzy connectives for both lower and upper ap-proximations. The experimental evaluation show that the datasets obtained with FRFS, compared with the raw datasets, achieve improvement of the ROC accuracy for some classification algorithms. Most important is that this is achieved with method that requires no information, threshold or domain information. The feature selected datasets achieve dimensionality reduction, not compromising the ROC accuracy.
Fuzzy rough sets, feature selection, dimensionality reduction, clas-sification