A Framework for Regularised Canonical Correlation Analysis by Alternating Least Squares

Large neuroimaging datasets often contain multiple views of each subject including imaging modalities and non-imaging data. Combining these high-dimensional views with Canonical Correlation Analysis (CCA) can provide valuable insights into their relationships [2, 4, 6]. In high-dimensional data, CCA requires regularisation. However many existing approaches to regularised CCA rely on restrictive assumptions [8] or approximate the CCA problem [7]. In this work, we introduce a flexible framework for regularising CCA by using alternating ‘blackbox’regularised least squares solvers, allowing users to choose their own solvers that fit the assumptions of their data or are optimised for it and demonstrate our framework using the Human Connectome Project (HCP) data.