Scalable Fuzzy Rough Set Reduct Computation Using Fuzzy Min–Max Neural Network Preprocessing

Abstract

A fuzzy rough set (FRS) is a hybridization of rough sets and fuzzy sets and provides a framework for reduct (feature subset selection) computation for hybrid decision systems. However, the existing FRS-based feature selection approaches are intractable for large decision systems due to the space complexity of the FRS methodology. We propose a novel fuzzy min–max neural network (FMNN)-FRS reduct computation approach utilizing the FMNN to enhance the scalability of FRS approaches. The FMNN provides a single pass epoch learning of arriving at granules of objects in the form of fuzzy hyperboxes for multiple decision classes. In the proposed approach, the FMNN model is used to reconstruct the object-based decision system into a fuzzy hyperbox-based interval-valued decision system. Then, a novel way of constructing the fuzzy discernibility matrix (FDM) from the interval-valued decision system is introduced. A fuzzy rough approximate reduct computation algorithm is developed with the induced FDM. The FMNN-FRS approach reduces the space complexity of FRS reduct computation significantly and results in enhanced scalability. Comparative experimental analysis has been done with the existing FRS reduct approaches on benchmark hybrid decision systems and established the relevance of the FMNN-FRS approach. The FMNN-FRS approach obtained the exact reduct in most of the datasets in much lesser computational time than existing FRS approaches while preserving similar classification accuracy. The FMNN-FRS method achieved enhanced scalability to such large decision systems, at which it is not possible to obtain reduct by existing FRS approaches.