Joint Embedding Learning and Low-Rank Approximation: A Framework for Incomplete Multiview Learning.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Real-world data often presents incomplete instances across multiple views.
• Incomplete Multiview Learning (IML) addresses this challenge by developing methods to handle partially available data.
• Existing IML techniques may lack a unified approach or efficient processing capabilities.

Purpose of the Study:

• To propose a novel framework, Joint Embedding Learning and Low-Rank Approximation (JELLA), for Incomplete Multiview Learning (IML).
• To unify existing IML methods and adapt complete multiview learning techniques for IML.
• To provide a foundation for developing new, efficient, and effective IML algorithms.

Main Methods:

• The JELLA framework approximates incomplete data using low-rank matrices.
• It learns a complete and common embedding through linear transformation.
• A new method, IML with Block-Diagonal Representation (IML-BDR), is proposed within the JELLA framework, utilizing block-diagonal structure for improved embedding.
• A convergent alternating iterative algorithm with successive over-relaxation is used for optimization.

Main Results:

• The JELLA framework successfully unifies several existing IML methods.
• It demonstrates the adaptability of complete multiview learning methods to IML.
• The proposed IML-BDR method, guided by the JELLA framework, shows effectiveness in clustering tasks.
• Experimental results validate the efficiency and accuracy improvements offered by the JELLA framework and IML-BDR.

Conclusions:

• The JELLA framework offers a unified and efficient approach to Incomplete Multiview Learning.
• It bridges the gap between complete multiview learning and IML, facilitating algorithm development.
• The IML-BDR method, a practical application of the JELLA framework, achieves more accurate clustering on various datasets.