[Paper Review] Canonical Correlation Analysis (CCA) Based Multi-View Learning: An Overview

Chenfeng Guo, Dongrui Wu|arXiv (Cornell University)|Jul 3, 2019

Face and Expression Recognition67 references32 citations

TL;DR

This paper surveys representative CCA-based multi-view learning approaches, highlighting linear, nonlinear, supervised, and deep extensions to fuse multiple views.

ABSTRACT

Multi-view learning (MVL) is a strategy for fusing data from different sources or subsets. Canonical correlation analysis (CCA) is very important in MVL, whose main idea is to map data from different views onto a common space with maximum correlation. Traditional CCA can only be used to calculate the linear correlation of two views. Besides, it is unsupervised and the label information is wasted. Many nonlinear, supervised, or generalized extensions have been proposed to overcome these limitations. However, to our knowledge, there is no overview for these approaches. This paper provides an overview of many representative CCA-based MVL approaches.

Motivation & Objective

• Motivate multi-view learning (MVL) as a way to fuse information from multiple views.
• Explain canonical correlation analysis (CCA) and its role in MVL as a shared subspace learner.
• Summarize representative CCA-based MVL methods and their key characteristics.
• Discuss extensions to handle more than two views, nonlinearity, sparsity, and supervision.

Proposed method

• Review traditional CCA and its limitations for multiple views and supervised tasks.
• Present GCCA and its variants for multiple views (SUMCOR, MAXVAR, SSQCOR, MINVAR, GENVAR).
• Describe nonlinear extensions (KCCA, LPCCA, DCCA, DCCAE, DisDCCA) and supervised/discriminative variants (DisCCA, MLDA, MULDA).
• Explain tensor/graph-based and high-order approaches (TCCA, RCCA, LPCCA).
• Outline deep learning integrations (DCCA, DisDCCA, DCCAE, DisDCCAE, VCCA, VCCA-private).
• Summarize optimization formulations (eigen decompositions, SVD, generalized eigenvalue problems) and training considerations.

Experimental results

Research questions

• RQ1 What are representative CCA-based approaches for two-view and multi-view data?
• RQ2 How do nonlinear, sparse, and supervised extensions address CCA limitations?
• RQ3 What are the connections between CCA-based methods and other MVL paradigms (discriminance, locality, deep learning)?

Key findings

• The paper provides a comprehensive overview of CCA-based MVL methods, including two-view and multi-view formulations.
• It covers linear, nonlinear, sparse, supervised, and deep learning extensions across multiple views.
• It contrasts kernel, locality-preserving, and discriminative variants with their underlying optimization problems.
• It highlights practical considerations like regularization and computational complexity for kernel and deep methods.
• It includes a comparative table (Table I) outlining main characteristics of representative approaches.