What are the cases when we should not use PCA for dimensionality reduction and what to use in such cases?
1
-
1$\begingroup$ Use PCA when it gets better results, or for visualizations.. $\endgroup$– HobbesCommented Jun 26, 2017 at 20:32
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
You should not use PCA if you only have categorical variables, and thus the distance function in PCA is invalid.
Correspondence analysis is a common alternative.
answered May 27, 2017 at 13:51
$\begingroup$
$\endgroup$
PCA is a linear transformation of your variables to a set of uncorrelated ones.
It is good when you want to remove redundancy (in a linear sense) in your data but it is bad if you want to uncover the "true cause" of the variables. Furthermore the causes can be non-linear in nature and the behavior of your data cannot be captured using a linear model.
Then you would stay away from PCA and consider e.g. non linear dimensionality reduction methods
