A novel approach to antinoise multi-granularity classification through graph-based feature selection

Abstract

This study proposes a novel anti-noise multi-granularity classification method that incorporates graph-based feature selection to effectively reduce computational complexity while enhancing classification accuracy and robustness. Traditional multi-granularity fuzzy rough set (MFRS) models are often highly sensitive to noisy samples, which hinders the acquisition of reliable knowledge. To address this issue, we design a heuristic feature selection algorithm within the framework of the weighted multi-granulation neighborhood-constrained fuzzy rough set (WMNcFRS) model. The algorithm first utilizes graph theory to evaluate inter-feature correlation and redundancy, enabling efficient partitioning of multi-granularity spaces. It then introduces a granularity-weighted strategy that ranks and prioritizes granularities based on approximation precision, thereby improving approximation capability. Finally, the feature selection strategy is formulated using the fuzzy dependency measure defined in the WMNcFRS model, which effectively suppresses noise interference, reduces the computational costs, and enhances the model's generalization ability. Extensive experiments on fifteen publicly available datasets demonstrate that, compared to six state-of-the-art algorithms, the proposed method exhibits superior robustness and classification performance, validating its effectiveness in practical applications.