Unsupervised Feature Selection Through Group Discovery

Modern scientific datasets are often high-dimensional and complex. When labels are unavailable or an unbiased analysis is desired, unsupervised feature selection can reveal meaningful structure, reduce noise, and promote interpretability. Most existing unsupervised feature selection methods evaluate features individually. Yet in many domains, informative signals arise from groups of related features that act jointly: adjacent pixels in images, functionally connected brain regions, and correlated financial indicators. Evaluating the importance of such features independently can therefore overlook group-level structure. Existing methods that incorporate feature groups often rely on predefined partitions or label supervision, which restricts their applicability. We address this limitation with GroupFS, an end-to-end, fully differentiable framework that simultaneously discovers latent feature groups and selects the most informative among them, without requiring predefined groups or labels. GroupFS combines differentiable feature-to-group assignments with stochastic group gates. It enforces Laplacian smoothness on both feature and sample graphs and applies a group-sparsity regularizer to learn a compact, structured representation that preserves the intrinsic geometry of the data while suppressing irrelevant features. Across nine benchmarks spanning image, tabular, and biological datasets, GroupFS consistently outperforms state-of-the-art unsupervised feature selection methods in downstream clustering tasks. In addition, GroupFS discovers coherent and interpretable feature groups that align with meaningful structure in the data.

M.Sc. student under the supervision of Prof. Ron Meir & Dr. Hadas Benisty.