How to minimize loss by groups for gradient boosting regression

Question

In my case I have 500 groups and 50k time series rows for them. They are inbalanced, so one group can have 5k rows and other 500 rows. I want to minimize the error for my gr