Multi-view subspace clustering with a consensus tensorized scaled simplex representation

Abstract

Listen

Translate article

Multi-view clustering has been at the forefront of research focused on identifying an ideal structure for the rational partitioning of sample points. Multi-view clustering faces two main challenges: identifying the optimal structure within each view and fusing these structures into a reliable consensus matrix. To overcome these obstacles, we introduce the Consensus Tensorized Scaled Simplex Representation (CTSSR). First, CTSSR uses the scaled simplex representation to reveal the structure within each view flexibly. Second, we apply a rank constraint strategy to ensure that the learned subspace structure embodies the optimal clustering structure. To fuse multiple views, we propose a normalized view weighting mechanism based on each view's reconstruction error, ensuring reasonable contributions to the consensus matrix. Finally, we employ tensor learning on weighted subspace views to explore high-order correlations and subtle differences between views, refining the exploration of inter-view relationships. CTSSR combines intra-view structure exploration with inter-view relationship analysis, resulting in an ideal consensus matrix. It shows excellent clustering performance on real-life datasets, significantly outperforming state-of-the-art methods in various experimental settings.