CEBP-HG: Hypergraph-regularized multi-subspace ensemble learning for robust unsupervised feature selection

Cong Guo

doi:10.1016/j.asoc.2026.115691

CEBP-HG: Hypergraph-regularized multi-subspace ensemble learning for robust unsupervised feature selection

2026-06-01

0

PRE

AI

黄昌勤 (Changqin Huang) *

X

Xiaodi Huang

DOI：10.1016/j.asoc.2026.115691

Original

Original request for help

Save

Abstract

En 中文

Unsupervised feature selection (UFS) is crucial for identifying discriminative features in high dimensional un labeled data, but existing methods remain limited by noise-prone pseudo-labels and the weak expressiveness of pairwise graph structures. To address these challenges, this paper proposes CEBP-HG, a novel unsuper vised feature selection method that combines multi-subspace cluster ensembling with hypergraph regularization. Specifically, CEBP-HG first partitions the original feature space into multiple random subspace sets and per forms independent clustering within each subspace. By ensembling these diverse clustering results, the method generates more stable and robust binary pseudo-labels that better approximate real-world label distributions. Furthermore, hypergraph regularization is employed to explore higher-order local instance relationships, en abling the capture of complex structural patterns beyond pairwise connections. Experimental evaluations on 13 datasets demonstrate that our algorithm consistently outperforms 15 state-of-the-art methods across 2 evaluation metrics. Moreover, our method achieves the highest overall performance ranking and exceeds the second-best approach by at least 2.71 ranks for each evaluation metric.

Keywords:

Feature selection

Unsupervised learning

Multi-subspace clustering

Ensemble learning

Hypergraph learning

CEBP-HG: Hypergraph-regularized multi-subspace ensemble learning for robust unsupervised feature selection

Abstract

Journal

Organization

Cited Papers