Unsupervised Feature Selection Through Group Discovery

The introduction of GroupFS marks a pivotal advancement in unsupervised feature selection, a critical area in high-dimensional learning tasks where labels are unavailable. Traditional methods often evaluate features independently, which can overlook the informative signals that arise from groups of related features. GroupFS addresses this limitation by jointly discovering latent feature groups and selecting the most informative ones without relying on predefined partitions or label supervision. By enforcing Laplacian smoothness on feature and sample graphs and applying a group sparsity regularizer, GroupFS learns a compact, structured representation. Its effectiveness is underscored by its performance across nine benchmarks, where it consistently outperforms state-of-the-art unsupervised feature selection methods. This capability is particularly relevant in diverse application areas such as images, tabular data, and biological datasets, highlighting the method's versatility and potenti…