I think we will be able to get away with average drop in gini!

It would be nice to support weighted observations so people don't have to rebalance their data sets. Unlike scikit-learn's we could actually specify what those weights mean and the range. haha.

If I remember correctly we have a shuffle function but there is one in the standard C++ library these days so we could avoid our own implementation that way. I just use this in my isolation forest so thought I would leave a note here.

int first = y[start] should be a float then do appropriate float comparison.

For broader adoption, I think we should make a drop in replacement for scikit-learn's. I.e., hyper parameters go in the constructor, fit(), predict(), predict_proba().

The one place we might want to deviate is the oob_score_ and feature_importances_. Probably should make these functions or at least get rid of the trailing underscore because these are public facing fields.