I used caret for a couple of years before making the switch to tidymodels this spring. I am getting comfortable with tidymodels, except it seems that the opportunities to integrate partial dependence plots are limited to single-predictor plots against the outcome variable.
I have seen that single-predictor plots are possible ( How to use `pdp:: partial()` in `tidymodels`), but I would also like to examine interaction, or two-predictor, plots for partial dependence, such as one can do with the pdp package with caret.
Should I return to caret and use pdp, or keep looking for a workaround with tidymodels? In addition, I understand that tidymodels processes data independently for each iteration within a k-fold cross validation. This avoids the data leakage that occurs when all data are processed at once and the training data can "peek" at the distribution of the testing data.
Does caret process data "within the fold" as well?