Profiling shows that the tree fitter spends a lot of time doing groupbys over the dimensions; worth testing if those groupbys go faster if the dimensions are cast to Categorical dtype. Might need some time, to test that this doesn't break anything, also categorical dims have non-intuitive groupby behavior.
Profiling shows that the tree fitter spends a lot of time doing groupbys over the dimensions; worth testing if those groupbys go faster if the dimensions are cast to Categorical dtype. Might need some time, to test that this doesn't break anything, also categorical dims have non-intuitive groupby behavior.