Removed a redundant field, intercept_steps, and a corresponding method to get the field. The field showed correct values except when link_function was "log" in which case it showed scaled values. This bug had no impact on model terms or predictions.
1.8.2
Estimating the intercept only once as the (weighted) average negative gradient, instead of updating it in each boosting step. This prevents predictors that are almost constants from competing with the intercept term.
1.8.1
- Estimating the intercept only once as the (weighted) average training response, instead of updating it in each boosting step. This prevents predictors that are almost constants from competing with the intercept term. - Fixed a minor bug related to min_observations_in_split.