Ya, learning causal structure from data is hard. The immediately connected nodes (parents or children) & colliders are minimal set of nodes that predict a variable assuming non-confounding. Then you typically remove colliders based on a bunch of independence tests. No method is ideal here!

I wouldnt touch variable selection with a 10 foot pole but LASSO kind of regression would be great for interpretability and for causal structure discovery kind of applications.

are there any recommendations regarding variable selection in GAMs?