Unsupervised feature selection (UFS) seeks a compact set of informative features without any class labels. Existing approaches typically rely on indirect structural cues such as similarity preservation, locality, sparsity, cluster geometry, or reconstruction error. This work reframes UFS as a representation learning problem on features rather than samples and introduces the Inverted Contrastive Learning framework (ICLFS).
ICLFS first transposes the data matrix so that each feature is represented by its sample‑profile vector. Multiple masked positive views are generated from this vector, together with a shuffled negative view. An InfoNCE‑based objective trains projector‑space embeddings to stay consistent across these structured perturbations.
Motivated by recent findings that cosine‑based and InfoNCE training affect embedding norms, the magnitude of the projector‑space embeddings is used as a saliency signal to rank features. The ranking is then refined by Laplacian‑Gated Ranking Correction, which suppresses locally redundant candidates while preserving salient ones.
Extensive experiments on twelve benchmark datasets show that ICLFS achieves the best clustering accuracy on ten datasets, outperforming both classical and neural baselines, and remains competitive on the remaining two. These results demonstrate that feature‑wise contrastive representation consistency offers a strong alternative to neighborhood, cluster, and reconstruction‑based UFS formulations.
Review