A Novel Feature Selection Method Based on Noise-Resistant Fuzzy Rough Sets
Abdelkhalek Hadrani, Karim Guennoun, Mohammed Wahbi · 2023
Feature selection or attribute reduction using Fuzzy Rough Sets (FRS) consists of reducing the global conditional attribute set of a given data set to a relevant feature subset of minimal or reduced size. This data preprocessing allows a given machine learning task to be conducted efficiently while preserving the semantics of the selected input variables in the considered data set. Many feature selection algorithms relied on fuzzy rough sets have been developed during the last two decades. Nevertheless, most of them are based on the first FRS concept proposed by Dubois and Prade in 1992 or on the classical FRS model, both of which are sensitive to class noise. Motivated by this fact, we propose in this paper a new feature selection method founded on one of the most noise-resistant FRS models, which is the FRS paradigm based on Automatic Noisy Sample Detection. Then, we give and analyze experimental results demonstrating the robustness and performance of our proposed method compared to four versions of the Fuzzy QuickReduct algorithm which are respectively built on the basis of the classical FRS paradigm and three most cited FRS models renowned for their robustness to noise.