Multiview Tensor Spectral Clustering via Co-Regularization.

Area of Science:

• Data Science
• Machine Learning
• Computer Science

Background:

• Multi-view clustering methods aim to integrate data from diverse sources by learning a consensus representation.
• Traditional methods struggle with high-dimensional data and capturing complex relationships within and across views.
• Existing affinity measures can collapse in high dimensions, hindering unified alignment.

Purpose of the Study:

• To develop a robust multi-view clustering approach for High-Dimensional Low-Sample Size (HDLSS) data.
• To address the limitations of traditional affinity measures in high-dimensional spaces.
• To propose a co-regularization framework for learning a fused consensus representation.

Main Methods:

• Encoding sample affinities using both dyadic and high-order measures for comprehensive spatial characterization.
• Learning a fused consensus representation by aligning multi-view low-dimensional data via co-regularization.
• Modeling fused representation learning through a high-order eigenvalue problem in manifold space, solved via manifold minimization.

Main Results:

• The proposed method effectively captures both intrinsic connections and complementary correlations within the data.
• Experiments on eight HDLSS datasets show superior performance compared to thirteen benchmark methods.
• The co-regularization approach successfully overcomes the limitations of traditional affinity measures in high dimensions.

Conclusions:

• The developed multi-view uniform clustering method via consensus representation co-regularization is effective for HDLSS data.
• The approach provides a robust way to handle data heterogeneity and high dimensionality in clustering.
• This work advances graph-based multi-view clustering by offering a more comprehensive and accurate representation learning technique.