Constrained ERM Learning of Canonical Correlation Analysis: A Least Squares Perspective.

Abstract

Canonical correlation analysis (CCA) is a useful tool in detecting the latent relationship between two sets of multivariate variables. In theoretical analysis of CCA, a regularization technique is utilized to investigate the consistency of its analysis. This letter addresses the consistency property of CCA from a least squares view. We construct a constrained empirical risk minimization framework of CCA and apply a two-stage randomized Kaczmarz method to solve it. In the first stage, we remove the noise, and in the second stage, we compute the canonical weight vectors. Rigorous theoretical consistency is addressed. The statistical consistency of this novel scenario is extended to the kernel version of it. Moreover, experiments on both synthetic and real-world data sets demonstrate the effectiveness and efficiency of the proposed algorithms.