Canonical Dependency Analysis based on Squared-loss Mutual Information

摘要

Canonical correlation analysis (CCA) is a classical technique to iteratively find projection directions for two sets of variables such that their correlation is maximized. In this paper, we propose an extension of CCA based on a squared-loss variant of mutual information. The proposed method, which we call least-squares canonical dependency analysis (LSCDA), has various useful properties, for example, it can capture higher-order correlations, it can simultaneously find multiple projection directions (i.e., subspaces), it does not involve density estimation, and it is equipped with a model selection strategy. We illustrate the usefulness of LSCDA through experiments.