A Granular-Ball Based Online Streaming Feature Selection Algorithm

Abstract

Online Streaming Feature Selection (OSFS) is a feature selection method to identify relevant features in real-time from high-dimensional, continuously generated data streams. Traditional methods require the pre-setting of static parameters and exhibit insufficient robustness. Therefore, these methods may fail to adapt effectively to evolving feature sets and are prone to noise interference. To overcome these limitations, this paper proposes an online streaming feature selection algorithm based on Granular Ball-Online Streaming Feature Selection (GB-OSFS). GB-OSFS adopts a granular ball approach in OSFS, shifting the focus from individual sample points to granular balls for calculating feature dependencies. Furthermore, we select granular balls with a purity of 1 as the evaluation metric for feature importance. As a result, this approach addresses the issues of traditional streaming feature selection methods being prone to noise interference and the necessity of pre-setting various hyperparameters. To validate the efficacy of GB-OSFS, a series of comprehensive empirical tests were conducted across ten datasets. The findings indicate that GB-OSFS surpasses six other OSFS algorithms in terms of prediction accuracy and robustness.