Fisher optimal splitting and missing values

[nickgiki]

In many situations missing values are unavoidable. XGBoost and LightGBM natively support them - for every candidate split they check how variance/entropy changes if the missing values go in the left/right node and pick the best.

Additionally, XGBoost and LightGBM support native categorical handling. One-hot encoding categorical features is suboptimal because, especially for features with high cardinality, one might need to grow very deep trees to find optimal splitting rules. Luckily there's Fisher's optimal splitting that reduces complexity from $O(2^{k-1})$ to $O(klogk)$ and at the same time provides more interpretable models (i.e. state in ('CA', 'NY')).

Is there any appetite to add these capabilities to the package? They will both really make a huge difference in building interpretable models!