TY - JOUR
T1 - Scalable Unaligned Multi-view k-Means
AU - Wang, Yichen
AU - Cao, Zhe
AU - Xin, Haonan
AU - Wang, Yuyang
AU - Wang, Rong
AU - Nie, Feiping
N1 - Publisher Copyright:
© 2026 IEEE. All rights reserved.
PY - 2026
Y1 - 2026
N2 - Conventional multi-view clustering methods typically assume exact sample correspondence across views. In real-world applications, this assumption is frequently violated due to heterogeneous sensing modalities and asynchronous data acquisition, leading to the view-unaligned problem (VuP). Although several representative methods have been proposed to address VuP, their computational cost remains a primary limitation, particularly in large-scale settings. To address this limitation, we propose Scalable Unaligned Multi-view k-Means (SUMK), an efficient framework that performs clustering without requiring predefined correspondences. SUMK formulates cross-view alignment from an entropy-regularized optimal transport (OT) perspective by converting the estimation of each view-wise permutation matrix into the computation of a corresponding transport plan, which is obtained by solving a view-wise OT problem. To improve scalability on large datasets, we further employ a sparsification mechanism for the OT solver, thereby substantially reducing computational and memory costs. Consequently, SUMK serves as a strong and scalable baseline for large-scale view-unaligned multi-view k-Means clustering. Despite its simplicity, SUMK is highly effective, delivering clustering performance that is competitive with or superior to state-of-the-art unaligned multi-view methods, while significantly improving efficiency and robustness.
AB - Conventional multi-view clustering methods typically assume exact sample correspondence across views. In real-world applications, this assumption is frequently violated due to heterogeneous sensing modalities and asynchronous data acquisition, leading to the view-unaligned problem (VuP). Although several representative methods have been proposed to address VuP, their computational cost remains a primary limitation, particularly in large-scale settings. To address this limitation, we propose Scalable Unaligned Multi-view k-Means (SUMK), an efficient framework that performs clustering without requiring predefined correspondences. SUMK formulates cross-view alignment from an entropy-regularized optimal transport (OT) perspective by converting the estimation of each view-wise permutation matrix into the computation of a corresponding transport plan, which is obtained by solving a view-wise OT problem. To improve scalability on large datasets, we further employ a sparsification mechanism for the OT solver, thereby substantially reducing computational and memory costs. Consequently, SUMK serves as a strong and scalable baseline for large-scale view-unaligned multi-view k-Means clustering. Despite its simplicity, SUMK is highly effective, delivering clustering performance that is competitive with or superior to state-of-the-art unaligned multi-view methods, while significantly improving efficiency and robustness.
KW - Cross-view Correspondence
KW - Multi-view Clustering
KW - Optimal Transport
KW - Unsupervised Learning
KW - View-unaligned Problem
UR - https://www.scopus.com/pages/publications/105052788491
U2 - 10.1109/TCSVT.2026.3741145
DO - 10.1109/TCSVT.2026.3741145
M3 - 文章
AN - SCOPUS:105052788491
SN - 1051-8215
JO - IEEE Transactions on Circuits and Systems for Video Technology
JF - IEEE Transactions on Circuits and Systems for Video Technology
ER -