Granular Ball-Based Noise-Resistant Fuzzy Multineighborhood Feature Selection via Label Enhancement and Feature Graph.

Abstract

Due to the increasing volume of multilabel data, interactions and complementarity among features are not fully explored in feature selection; the descriptive differences of labels to samples are frequently overlooked, and the abundant features and noise adversely affect classification efficacy. To solve these challenges, this article constructs a granular-ball-based, noise-resistant fuzzy multineighborhood feature selection scheme leveraging label enhancement and a feature graph. First, an overall similarity between samples is developed via label Jaccard similarity and the Pearson correlation coefficient, thereby grouping similar samples into the same granular ball. Second, a multineighborhood radius is designed via the inherent properties and distributions of features, resulting in adaptive fuzzy multineighborhood granules. A granular-ball-based, noise-resistant fuzzy multineighborhood rough set model will be constructed by combining multineighborhood fuzzy decisions. Third, by using the similarity of samples in a two-space fusion of both features and labels, the discriminant of samples under different labels is obtained, and label enhancement is achieved via fuzzy multineighborhood granules. Uncertainty measures are derived, and the relevance, association, redundancy, complementarity, and interactivity in multilabel data are studied. Finally, the multiple correlation relationships between features and labels are utilized to develop a weighted feature graph, and the significance of feature vertices can be evaluated to guide multilabel feature selection. Extensive experiments on 14 datasets demonstrate that our scheme outperforms existing state-of-the-art methods.