Noise-aware and correlation analysis-based for fuzzy-rough feature selection

Abstract

Feature selection has gained significant attention, with a focus on removing redundant or irrelevant features to improve subsequent machine learning tasks. The fuzzy rough set is extensively utilized for feature selection because of its ability to manage uncertainty. The fuzzy dependency-based approach represents an effective branch within the domain of fuzzy rough set models. However, it has been revealed that fuzzy dependency function is not robust to noise data while the noise objects often occur. Furthermore, as the size of dataset increases, the computational complexity of evaluating metrics also tends to increase due to potentially unnecessary computational effort. Hence, to address both of these research issues simultaneously, we propose a fuzzy rough correlation-based feature selection algorithm, denoted as FRCB. The introduction of neighborhood radius constraints and granularity conditions aims to interpret and mitigate the impact of noisy data by partitioning objects into a Core-Boundary-Outlier structure. A more robust hybrid fuzzy relation and correlation-based fuzzy rough dependency is proposed by considering the representativeness and attractiveness of the core to other objects within the internal class, as well as the impacts on other objects in external classes. To minimize processing overhead and time complexity, redundant features are identified through correlation analysis, which involves evaluating, ranking, and grouping before proceeding to the next iteration. The experiments were conducted using eleven datasets. The analysis of the noise curves and the raw data curve indicates that the proposed correlation-based fuzzy rough dependency is more robust than other benchmarking methods. Moreover, the results of the classification and running time comparisons indicate that the FRCB algorithm proposed in this study exhibits superior performance compared to the baseline methods.