Beyond Low-Rank Representations: Orthogonal clustering basis reconstruction with optimized graph structure for...

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Low-Rank Representation (LRR) is a powerful paradigm for multi-view spectral clustering.
• LRR encodes local graph structures into a low-rank self-expressive similarity for improved graph partitioning.
• Existing LRR methods effectively capture multi-view data but can be further optimized.

Purpose of the Study:

• To fundamentally revisit and reinterpret Low-Rank Representation (LRR) from a novel perspective.
• To propose a new technique based on latent clustered orthogonal projection for multi-view spectral clustering.
• To enhance the learning of cluster structures and optimize local graph structures within multi-view data.

Main Methods:

• Decomposing LRR into a latent clustered orthogonal representation using low-rank matrix factorization.
• Simultaneously learning orthogonal clustered representations and optimized local graph structures for each view.
• Achieving multi-view consensus by ensuring learned representations and graph structures have the same magnitude.

Main Results:

• The proposed method decomposes LRR into a more flexible latent clustered orthogonal representation.
• It effectively learns both orthogonal clustered representations and optimized local graph structures.
• Experimental results demonstrate superior performance compared to state-of-the-art LRR models on multi-view datasets.

Conclusions:

• The novel perspective on LRR as latent clustered orthogonal projection offers significant advantages for multi-view spectral clustering.
• The proposed technique provides a more flexible and effective approach to encoding cluster structures and graph properties.
• This method achieves superior performance, highlighting its potential for advanced data analysis and clustering tasks.