Online Canonical Correlation Analysis via Rayleigh-Ritz Projections

Vassilis Kalantzis, Panagiotis A. Traganitis, Charilaos I. Kanatsoulis

Abstract: Canonical correlation analysis (CCA) is a classical statistical tool that enables processing of multiview data in a plethora signal processing, machine learning and data mining applications, by identifying a common linear subspace from the available data views. Most algorithms tackling the CCA task require all the data per view to be available. Nevertheless, in many cases, data are not available in batch form and may arrive in streaming fashion. This work puts forth a novel, computationally efficient, projection based method for identifying and updating the common subspace on-the-fly, as new data arrive, while retaining its' fidelity. Preliminary numerical tests, on synthetic and real data benchmarks, showcase the potential of the proposed method.