Weak learners for XGBoost with Tweedie distribution Could you please explain what are the standard weak learners for XGBoost 
when the objective parameter equals reg:tweedie? 
Are they  GLMs (with Tweedie distribution of dependent variable) on 
all possible subsets of the set of my predictors? Are GLMs with interactions of the form x_i*x_j also included? 
 A: The standard base learners when selecting the use of reg:tweedie are tree-based models, i.e. gbtree. We can use a linear model as a base learner but it is not by default. When we set reg:tweedie we effectively a Tweedie distribution's negative log-likelihood as your objective function and we do not directly affect the base learner. (To be a bit more precise: we actually use the gradient and the hessian of Tweedie log-likelihood in the objective function specification; the negative log-likelihood of the Tweedie regression is used as the evaluation metric but the evaluation metric can be changed to something like MAE and still used reg:tweedie as our objective function. The code implementing these computations for the Tweedie objective can be found here.)
Finally, no, interactions between variables are not directly considered. If we think a particular interaction should be considered we should form it manually. If interaction effects occur, they are expected to be captured through the boosting procedure.
