Large-scale structured subspace clustering

Mo Chen

doi:10.1007/s10618-026-01268-x

Return

Large-scale structured subspace clustering

2026-09-21

PRE

DOI：10.1007/s10618-026-01268-x

Original

Original request for help

Save

Abstract

En 中文

Subspace clustering has become an essential technique in data mining and pattern discovery for revealing low-dimensional structures embedded within high-dimensional data. By modeling data as a union of multiple linear or affine subspaces, it effectively captures complex latent relationships among samples. However, existing methods still face two major challenges: (1) scalable representations may not adequately preserve intrinsic structural relationships under complex data distributions; (2) global self-representation models incur prohibitive computational costs as data scale increases. To address these limitations, we propose a Large-scale Structured Subspace Clustering (LSSC) framework that integrates anchor-based reconstruction, locality-aware coefficient initialization, distance-weighted structure regularization, and coefficient regularization to learn a compact nonnegative sample-to-anchor representation. The proposed method achieves strong clustering performance while avoiding the construction of a full sample-to-sample affinity matrix, thereby providing favorable scalability on large datasets. Extensive experiments on nine benchmark datasets demonstrate that LSSC achieves the best overall average rank across ACC, NMI, Purity, and ARI among the compared methods that completed all nine datasets, while maintaining competitive clustering performance and scalability on large-scale datasets. These findings characterize both the practical scalability of LSSC and its sensitivity to highly imbalanced data distributions, thereby clarifying its empirical strengths and limitations.

Keywords:

Anchor-based modeling

Representation learning

Affinity graph

Structure-aware

Manifold regularization