Feature Selection and Classification Based on Directed Fuzzy Rough Sets

Abstract

Fuzzy rough sets have made considerable strides within the domain of machine learning and data mining and served as a valuable tool for feature selection. However, traditional models face challenges in computing fuzzy similarity relations. They oversimplify the treatment of diverse samples by assuming that they exist in the same class space, ignoring their labels and distribution information. Consequently, difficulties arise when dealing with data that exhibit considerable distribution variations across classes. To address this issue, this study proposes a directed fuzzy rough set model that better captures the inherent uncertainty in sample distribution compared with traditional models. In this model, class-subspace distribution information is seamlessly integrated into directed fuzzy binary relations. Furthermore, fuzzy rough approximation operators are redefined to accurately capture the uncertainty associated with class distribution, facilitating a comprehensive analysis of relevant properties concerning decision approximations for samples. Building on this background, a heuristic algorithm for feature selection and a K-nearest neighbor reduction classifier are developed. Comparative experiments with top-tier algorithms showcase the outstanding performance of our proposed model. This study provides a robust framework for addressing intricate machine learning and pattern recognition tasks.