OSHO-CCA: Orthogonal and Scalable High-Order Canonical Correlation Analysis

TMLR Paper6714 Authors

Abstract: Canonical Correlation Analysis (CCA) is a classical technique for learning shared representations from two views of data by maximizing the correlation between the resulting representations. Existing extensions to more than two views either maximize pairwise correlations, sacrificing higher-order structure, or model high-order interactions at the expense of orthogonality and scalability. In this paper, we propose OSHO-CCA, a novel method for Orthogonal and Scalable High-Order CCA that jointly addresses all three desiderata: (1) it captures high-order dependencies across views, (2) enforces orthogonality among projected features to ensure decorrelated embeddings, and (3) scales efficiently with the number of views. We further introduce a new evaluation metric for Total Canonical Correlation (TCC) that generalizes traditional two-view CCA metrics to the multiview setting. Experiments on real and synthetic datasets demonstrate that OSHO‑CCA outperforms existing methods in both correlation maximization and downstream classification tasks, while maintaining scalability and orthogonality even in challenging multiview scenarios.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Jaakko_Peltonen1

Submission Number: 6714