Is there any other role of using PCA in the feature level fusion based on Canonical correlation analysis or only used for dimensionality reduction?
Usually PCA is used as a dimensionality reduction step before CCA for fusion in most applications. However, care must be taken as PCA retains components that have most variance in one data set and therefore, these might not be the same components that have most correlation between two sets. A joint PCA-CCA approach, thus, is a better solution. However, for some applications, PCA as a dimensionality reduction step might not always work and therefore other techniques such as sparse CCA might provide more interpretable results.
Thank you Tanuj Hasija
according to your explanation, is it possible to joint any other dimentionality reduction techniqe with CCA for fusion purpose?
Yes, for instance sparse CCA. However, it comes at a cost of more complexity and computation time compared to a PCA pre-processing step.
So the reason of using PCA before CCA, is the low performance of the CCA technique with high dimensional data. Therefore, the PCA used to reduce the dimensionality before the fusion process
- Badges
- Science method