I didn't have enough patience to wait for Max to make a tidymodels/principles repo, so I made this outline and filled in the smattering of notes I have.
Things I think we need to understand and treat as separate throughout:
- programmatic versus interactive modelling
- end goal of inference versus end goal of prediction
- a single fit versus a set of fits
Other things to do:
- decide on some canonical modelling examples. proposals:
- LASSO for a model specification with structure hyperparameter space
- random forest for a model specification with unstructured hyperparameter space
Another potential topic: wrapping and working with existing modelling packages