Unpaired Canonical Correlation Analysis
This presentation introduces Unpaired Canonical Correlation Analysis (UCCA), a method that learns shared representations across two views without any paired training examples. The paper reframes unpaired multiview learning as a structural matching problem, showing that finding the permutation that maximizes canonical correlation is equivalent to solving a graph-matching problem between the views' geometric kernels. UCCA demonstrates that abundant unpaired data alone provides a strong enough alignment signal to recover canonical correlations approaching supervised methods on real-world benchmarks spanning handwritten digits, multi-omics, and image-text data.Script
Most machine learning methods for multiview data rely on a simple assumption: they know which observation in one view corresponds to which observation in the other. But what happens when that correspondence is completely unknown?
The researchers tackle Unpaired Canonical Correlation Analysis, where two views are observed independently with no correspondence available during training. This is fundamentally different from domain adaptation or manifold alignment, because the goal is not just distributional similarity but recovery of the specific permutation that maximizes canonical correlation.
The paper's core insight is a reduction. Under a technical relaxation, the problem of finding the permutation that best aligns canonical correlations becomes equivalent to a Quadratic Assignment Problem on the views' linear kernel matrices. Instead of searching over joint transformations and permutations, the method matches the geometric structure encoded in each view's within-sample similarity.
UCCA operationalizes this reduction through anchor-based matching. Each view is independently clustered to extract representative centroids, the Quadratic Assignment Problem is solved on these anchors, and standard Canonical Correlation Analysis is then fitted to the resulting pseudo-pairs. This anchor strategy reduces combinatorial cost while preserving the manifold geometry necessary for alignment.
Across six real-world configurations, UCCA recovers total correlations of 1.957 on handwritten digits and 1.755 on single-cell multi-omics data, approaching paired upper bounds of 1.990 and 1.850. The gap is larger on image-text tasks, where modality-specific variation and semantic ambiguity reduce the strength of geometric correspondence. The method consistently exceeds all unpaired baselines, and downstream cross-view classification confirms that it preserves shared cluster structure even without instance-level supervision.
Unpaired Canonical Correlation Analysis demonstrates that abundant unpaired data alone can provide a strong alignment signal when the views retain identifiable shared geometry. The work opens a technically grounded path for multiview learning without correspondence, and you can explore the full method and create your own video explanations at EmergentMind.com.
