Open ChungsooKim opened 2 years ago
we can input model agnostic importances (interpretability) metrics like SHAP, LIME, or permutation importance in the VarImp column. but torch in R doesn't seem to have tools calculating those metrics yet.
I thought about SHAP, but it would take tons of time to screen all covariates...
Yes, that's right. we will use millions of covariates from millions of patients. Then we can store all the original values of the weights for each layer in a separate column. However, if we prepare the code for calculating the importance metric in advance, users can only calculate the metric if they want (optionally). I would try to get agnostic codes for the torch.
As @jreps mentioned prior meeting (21-10-04),
I also found that all variable importance of the fitted model is zero.
However, I also figured out the final model from the fitDeepNNTorch function still had values of predictors as weights and biases in each layer. so we need to discuss how to deal with these values for model reproducibility