Diverse Non-Negative Matrix Factorization for Multiview Data Representation.

Area of Science:

• Machine Learning
• Data Mining
• Computational Science

Background:

• Non-negative Matrix Factorization (NMF) is a key technique for parts-based data representation.
• Real-world data often presents multiple views and high dimensionality, necessitating methods that exploit diversity and efficiency.
• Existing NMF methods may not fully leverage multiview information or handle high-dimensional data efficiently.

Purpose of the Study:

• To develop novel Non-negative Matrix Factorization (NMF) approaches that enhance data representation by incorporating diversity from multiple views.
• To address the challenges of high-dimensional data by proposing efficient low-dimensional representation learning methods.
• To introduce a diversity term that reduces redundancy and improves the comprehensiveness of multiview representations.

Main Methods:

• Proposed a diverse NMF (DiNMF) approach incorporating a novel diversity term to enhance multiview representations and reduce redundancy.
• Developed a locality preserved DiNMF (LP-DiNMF) to ensure diversity while preserving local data geometry.
• Derived efficient iterative updating algorithms for both DiNMF and LP-DiNMF with convergence proofs.

Main Results:

• Demonstrated that DiNMF enhances diversity and reduces redundancy in multiview representations.
• Showcased LP-DiNMF's ability to maintain data structure while ensuring multiview diversity.
• Experiments confirmed the efficiency and accuracy of DiNMF and LP-DiNMF compared to state-of-the-art methods on synthetic and real-world datasets.

Conclusions:

• The proposed diversity term significantly improves NMF by enhancing multiview representation quality.
• DiNMF and LP-DiNMF offer efficient and accurate solutions for learning low-dimensional representations from high-dimensional, multiview data.
• The methods provide a competitive advantage in data analysis by effectively exploiting data diversity.