Two view learning: SVM-2K, Theory and Practice

Abstract

Kernel methods make it relatively easy to define complex highdimensional
\nfeature spaces. This raises the question of how we can
\nidentify the relevant subspaces for a particular learning task. When two
\nviews of the same phenomenon are available kernel Canonical Correlation
\nAnalysis (KCCA) has been shown to be an effective preprocessing
\nstep that can improve the performance of classification algorithms such
\nas the Support Vector Machine (SVM). This paper takes this observation
\nto its logical conclusion and proposes a method that combines this
\ntwo stage learning (KCCA followed by SVM) into a single optimisation
\ntermed SVM-2K. We present both experimental and theoretical analysis
\nof the approach showing encouraging results and insights.