Uncertainty-Aware Online Sparse Streaming Feature Selection Via Particle Swarm Optimization

High-dimensional streaming data applications commonly employ online streaming feature selection (OSFS) techniques. However, practical implementations often encounter data incompleteness issues stemming from equipment malfunctions and technical limitations. Online sparse streaming feature selection (OS^2FS) addresses this challenge through missing data imputation using latent factor analysis. Current OS^2FS approaches exhibit significant limitations in handling uncertain feature-label relationships, resulting in rigid models with suboptimal performance. To overcome these limitations, this paper introduces an original uncertainty-aware online sparse streaming feature selection via particle swarm optimization (UOS^2FS) framework that incorporates two key innovations: (1) particle swarm optimization-guided supervision to mitigate uncertainty impacts, and (2) three-way decision theory integration for handling feature fuzziness in supervised settings. Comprehensive evaluations across six Practical datasets confirm the superior performance of UOS^2FS over existing OSFS and OS^2FS methods, consistently achieving higher accuracy through more effective feature subset selection.