Noise-Tolerant Feature Selections Based on Two-Type Weight-Fuzzy Granulations and Three-View Uncertainty Measures
Xin Xie, Xianyong Zhang, Xiaoling Yang, Jilin Yang · IEEE Transactions on Knowledge and Data Engineering · 2025
Noise-tolerant feature selections are valuable for data learning; they can resort to efficient fuzzy granulations and uncertainty measures, and a fundamental model concerns weighted kernel fuzzy rough sets (WKFRSs) which consider data distributions and uncertainty. In terms of current WKFRSs, fuzzy granulations adopt k-nearest neighbors for weighted optimization, while uncertainty measures consider single algebraic and informational views; corresponding feature selection algorithms have made achievements of noisy processing, but still exist advancement space from granulation deepening and measurement reinforcement. In this paper embracing WKFRSs, two-type weight-fuzzy granulations are defined by using self-adapting radius neighborhoods, three-view uncertainty measures are comprehensively constructed from uncertainty mechanisms, so 2 × (1 + 1 + 2) = 8 heuristic algorithms of feature selections are systematically established for better noise-aware learning. At first, two improved factors of local density and boundary influence are proposed by general neighborhood characterization and statistical radius determination, and thus two sample weights emerge to adjust Gaussian-kernel fuzzy relations to induce two weight-fuzzy granulations. Then, the fuzzy precision and fuzzycomplementary mutual information are respectively proposed from algebraic and informational views, and the two are combined into two fused measures via arithmetic and geometric means. Furthermore, the above two-type granulations and threeview measures two-dimensionally generate 2×(1+1+2) = 8 new heuristic selection algorithms via feature significances. Finally by data experiments, constructional fuzzy granulations, uncertainty measures, feature selections are validated to have anti-noise characteristics and corresponding robustness, while new selection algorithms acquire better performances of classification learning than multiple contrast algorithms.