Adopt the linear_tree feature of lightGBM to h2o.xgboost
Currently, lightGBM has a 'linear_tree' feature, which enables users to build piecewise linear gradient boosting tree. Is it possible to adopt it in the h2o.xgboost()? Sometimes this feature significantly improves the predictive power of the final model.
Currently this is not possible with xgboost, once this is added to xgboost we will be happy to add support to h2o’s xgboost
Thank you for the update.
Xgboost does not have this option but LightGBM does. Since h2o emulates LightGBM by changing some hyperparameters of h2o.xgboost(), I am wondering whether this unique feature of LightGBM can be included as well.
In addition, the monotone_constraints of h2o.xgboost() does not work with gblinear; could you modify the function to make it so? The signs of coefficients of the final model are critical.
Appreciate your help.
AFAIK xgboost has no such option, it only has the gblinear updater, but this results in non-tree based linear models
please take a look