Online Stable Streaming Feature Selection via Multiscale Bidirectional Redundancy and Multicenter Dynamic Feature Clustering

By processing dynamically arriving features, online streaming feature selection proves highly suitable for real-world applications, leading to growing attention. However, existing online streaming feature selection methods are generally based on a single-scale assumption and ignore that the same feature may exhibit different discriminative abilities at multiple scales. In addition, these methods pay insufficient attention to the stability of feature selection results. To address these issues, we propose an online stable streaming feature selection method via multiscale bidirectional redundancy and multicenter dynamic feature clustering, named MFC-OGSFS. First, the concept of multiscale information systems is introduced into online streaming feature selection, and a scale fusion weight based on consistency degree is proposed. Second, a new metric, called the multiscale fuzzy Gini granularity, is proposed to quantify feature uncertainty in multiscale fuzzy information systems. Third, a multicenter updating mechanism and a weighted fusion strategy are designed to improve the stability of feature selection. Based on the multiscale fuzzy Gini granularity, we further propose the concepts of bidirectional redundancy and multicenter supplementation, which construct a three-stage online stable streaming feature selection method. Finally, extensive experiments are conducted on multiple real-world datasets to compare the proposed method with ten state-of-the-art algorithms in terms of classification accuracy, stability, and the number of selected features. Experimental results show that MFC-OGSFS simultaneously achieves the optimal classification accuracy and feature selection stability while selecting a small number of features.