Answers are grounded in this paper and cite their sources.
Unsupervised feature weighting relies on sample structure to assess feature importance. However, irrelevant features and noise can distort the structural reference, while feature redundancy can cause repeated weighting of similar information. We propose an unsupervised feature-weighting method based on a graph heat kernel and the normalized Hilbert–Schmidt independence criterion (nHSIC), termed GHK-nHSIC-FW, which constructs a label-free structural target using heat diffusion on a nearest-neighbor graph. The method uses nHSIC to measure feature–target alignment and feature redundancy and learns continuous weights with convex quadratic optimization under nonnegative simplex constraints to adjust feature contributions to the weighted Euclidean distance. Experiments on eight public datasets demonstrate that our proposed method can achieve the highest average clustering accuracy (ACC) and normalized mutual information (NMI) among the compared methods, while ablation and subsampling analyses assess component effectiveness and feature importance stability. Synthetic experiments with known informative features examine sample contamination, irrelevant feature injection, correlated feature copies, and diffusion-scale effects under graph perturbations. The results show strong relative performance under sample and irrelevant feature contamination and demonstrate how redundancy control helps retain complementary informative contributions. These results support graph-guided, redundancy-aware feature weighting for unsupervised clustering.
No takes yet. Share an insight, caveat, or question.
Zou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: