Complete Model Selection in Multiset Canonical Correlation Analysis

Traditional model-order selection for canonical correlation analysis infers latent correlations between two sets of noisy data. In this scenario it is enough to count the number of correlated signals, and thus the model order is a scalar. When the problem is generalized to a collection of three or more data sets, signals can demonstrate correlation between all sets or some subset, and one number cannot completely describe the correlation structure. We present a method for estimating multiset correlation structure that combines source extraction in the style of joint blind source separation with pairwise model order selection. The result is a general technique that describes the complete correlation structure of the collection.