Efficient Canonical Correlation Analysis with Sparsity
Elena Tuzhilina, PhD
Assistant Professor聽Department of Statistical Sciences, University of Toronto
WHEN: Wednesday, March 18, 2026, from 3:30 to 4:30 p.m.
WHERE: Hybrid | 2001 麻豆传媒网站 College Avenue, Rm 1140;
NOTE:聽Elena Tuzhilina will be presenting in-person at SPGH聽
Abstract
Canonical Correlation Analysis (CCA) is a fundamental multivariate method for measuring associations between two datasets, with applications in genomics, neuroimaging, public health, and machine learning. In high-dimensional settings, however, classical CCA breaks down, and existing sparse approaches often require a difficult trade-off between computational efficiency and statistical guarantees. We introduce ECCAR, a fast and provably consistent sparse CCA algorithm that resolves this tension. By reformulating CCA as a high-dimensional reduced-rank regression problem, we obtain consistent estimators with high-probability error bounds while avoiding computationally intensive procedures such as Fantope projections. The resulting method is scalable, projection-free, and substantially faster than competing approaches. We validate ECCAR through extensive simulations and demonstrate its practical utility on real-world data, including an Alcohol Use Disorder and the Autism Brain Imaging Data Exchange datasets, where it uncovers reliable and interpretable multivariate associations.
Speaker Bio
Elena Tuzhilina is an Assistant Professor in the Department of Statistical Sciences at the University of Toronto. She received her Specialist鈥檚 degree in Mathematics from Moscow State University in 2015 and her PhD in Statistics from Stanford University in 2022 under the supervision of Professor Trevor Hastie, where she focused on developing scalable statistical methods for structured and high-dimensional data with applications in biology and medicine. Her research lies in applied statistics, with methodological contributions centered on dimension reduction and latent space modeling. She develops statistical methods that facilitate the analysis of complex biomedical datasets, with applications in areas such as neuroscience, conformation reconstruction, and single-cell data analysis..