Opened 9 months ago
Handle missing values (i.e. NaN) correctly when training GBT models
|Reported by:||gkronber||Owned by:||gkronber|
Right now GBT training only uses comparison operators and assumes that there are no missing values in the input variables. This is not really correct as NaN values are assigned to either the left or the right subtree. Instead those observations should be ignored in the training phase.
This ticket is related to #2612.