R⁵-FRQR: Robust and Interaction-Aware Feature Selection Using Fuzzy Rough Sets

Traditional fuzzy rough quick reduct (FRQR) feature selection methods often rely on greedy, marginal relevance evaluation, which overlooks feature interactions. This leads to instability during resampling and consequently performs poorly under domain shifts. To overcome these limitations, this paper introduces R⁵-FRQR, a fuzzy rough set-based feature selection framework that integrates five properties: reliability, relational awareness, robustness, reductiveness, and resilience (R ⁵ ). R⁵-FRQR incorporates an interaction-aware selection mechanism that accounts for both feature redundancy and pairwise contribution to dependency scoring. A granular similarity approximation strategy is introduced to reduce computational cost while preserving selection quality. Moreover, feature stability is reinforced through bootstrap-based frequency scoring. A domain-adaptive dependency metric is proposed to maintain discriminative power across distribution shifts. Experimental results on 16 benchmark datasets demonstrate that R⁵-FRQR consistently achieves higher performance across evaluation metrics compared to classical and state-of-the-art feature selection methods. The framework provides a scalable and interpretable solution for robust feature selection in high-dimensional, noisy, and dynamic data environments.