B1337
Title: Matrix-variate canonical correlation analysis with a network neuroscience application
Authors: Daniel Kessler - University of North Carolina at Chapel Hill (United States) [presenting]
Liza Levina - University of Michigan (United States)
Abstract: The extension of canonical correlation analysis (CCA) is considered to be the matrix-variate setting, where one or both of the random vectors of classical CCA are replaced by random matrices. The goal remains the identification of pairs of linear functions that transform the data into maximally correlated canonical variates. The matrix-specific structure is exploited by seeking low-rank representations through the use of a nuclear norm penalty. Although generally applicable to matrix-variate data, this approach is motivated by applications in network neuroscience, where the matrix-variate data is a participant-specific connectivity matrix of spatial correlations. When applied to network data, these low-rank canonical directions can be understood as seeking latent network structure. It is shown in synthetic data that the approach is effective at recovering low-rank signals even in noisy cases with relatively few observations, and the method is applied to human neuroimaging data.
