An extended Oja process for streaming canonical analysis
Jean-Marie Monnez · Communication in Statistics- Theory and Methods · 2025
Canonical components of the canonical analysis of two random vectors are collinear with principal components of a PCA of the multidimensional linear regression function of one vector with respect to the other or projected PCA. In the context of streaming data, we define processes to estimate online in parallel this regression function and components of the canonical analysis, possibly taking into account at each step all the data up to this step to have a faster convergence and using an extended Oja process. We deal with the cases of canonical correlation analysis, factorial correspondence analysis and factorial discriminant analysis.