Flexible Multi-View Dimensionality Co-Reduction.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Dimensionality reduction is crucial for handling high-dimensional data.
• Unsupervised learning methods are needed for data with multiple, potentially disparate, views.
• Existing methods often struggle with unbalanced dimensionality across multiple data views.

Purpose of the Study:

• To propose a novel unsupervised dimensionality reduction method for multi-view data.
• To exploit the complementarity and consistency between multiple data views.
• To develop a method that handles unbalanced dimensionalities and supports out-of-sample data projection.

Main Methods:

• Developed Multi-view Dimensionality co-Reduction (MvDR) algorithm.
• Employed kernel matching with Hilbert-Schmidt Independence Criterion (HSIC) to enhance inter-view correlations.
• Ensured locality within views and consistency across views via joint optimization.

Main Results:

• MvDR successfully reduces dimensionality while preserving essential data structures.
• The method demonstrates superior performance in clustering and recognition tasks compared to state-of-the-art approaches.
• Individual low-dimensional projections are generated, enabling out-of-sample data application.

Conclusions:

• Multi-view Dimensionality co-Reduction effectively integrates information from multiple data views.
• The proposed kernel matching approach robustly handles view complementarity and unbalanced dimensionalities.
• The method offers a significant advancement for multi-view unsupervised learning applications.