Generalization of CCA via Spectral Embedding
Jagadeesh Jagarlamudi · 2011
Given a multi-view data, Canonical Correlation Analysis (CCA) [3] is a technique to find the projection directions in each view so that the observations when projected along these directions are maximally aligned. Let X (d1×n) and Y (d2×n)be the representation of data in both the views, then CCA finds the projection directions a and b such that: argmax a,b aXY b √ aTXXTa √ bTY Y Tb This objective function can be re-written as: