Hi everyone,
I am trying to conduct a PCE-based sensitivity analysis for feature selection of a dataset. However, the LOO error of the trained PCE model appears to be unacceptably high, approximately 0.4. I suspect there’s room for improvement, considering that the dataset I’m using isn’t directly derived from a computational model. In this case, while I understand that the Sobol’ indices may not be accurate, but in terms of getting the importance ranking of the inputs, is it still okay to use the sensitivity analysis results based on the trained PCE model?
Thanks in advance for any insights.
Best,
Hongzhou