Multiview Clustering by Consensus Spectral Rotation Fusion.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multiview clustering (MVC) leverages complementary information from multiple data sources.
• Existing MVC methods often fuse raw data, leading to performance issues due to redundant information.
• Graph learning methods in MVC can be limited by specific graph constructions and high computational complexity (O(n³)).

Purpose of the Study:

• To propose a novel Consensus Spectral Rotation Fusion (CSRF) method for multiview clustering.
• To address limitations of existing MVC techniques, including raw data fusion and computational inefficiency.
• To learn a fused affinity matrix at the spectral embedding feature level for improved MVC.

Main Methods:

• Developed a CSRF model to learn a consensus low-dimensional embedding by exploring cross-view complementary and consistent information.
• Implemented an alternating iterative optimization algorithm with reduced computational complexity (O(n²)) per iteration.
• Integrated a sparsity policy for graph construction and extended CSRF for incomplete MVC scenarios.

Main Results:

• The proposed CSRF method effectively fuses multiview information at the spectral embedding level.
• Achieved a lower computational complexity compared to traditional matrix inversion or eigenvalue decomposition methods.
• Demonstrated effectiveness and efficiency through extensive experiments on various multiview datasets.

Conclusions:

• CSRF offers an effective and efficient approach to multiview clustering by operating at the spectral embedding feature level.
• The method overcomes limitations of raw data fusion and high computational costs associated with existing techniques.
• CSRF shows promise for handling incomplete multiview data, expanding its practical applicability.