CODEWITHSUNDEEP
rxgboostmlr
Andrea
Andrea

Reputation: 1

is there a way to set the "base margin" for XgBoost in the MLR framework?

I am trying to fit an XgBoost model within the MLR framework. While the framework is fairly well documented, there are some specifics of the XgBoost library that I cannot replicate within MLR, one in particular being the "base margin". In XgBoost library I would just set :

db_xgbmatrix <- xgb.DMatrix(db)
setinfo(db_xgbmatrix , "base_margin", margin)

and then I can just go on with the training of the model; whereas, in MLR, once I create the task and the learner: 

tsk      <- makeRegrTask(data =  db, target = target_var)
lrn <- makeLearner("regr.xgboost", predict.type = "response", eta = 0.1,
                   max_depth = 8, min_child_weight = 20,
                   subsample = 0.75, colsample_bytree = 0.75,
                   nrounds = 100, nthread = cl_n, objective="count:poisson")

I'm not quite sure of where the base margin should be set. Any ideas? Is that feature implemented, and hidden somewhere? Thank you all in advance

Upvotes: 0

Views: 1554

Answers (2)

John
John

Reputation: 609

This is finally supported in mlr3 and mlr3learners for xgboost (currently using the latest github versions).

See example here. Setting offset in the task means also that it is applied automatically during internal tuning/validation of xgboost as well.

Upvotes: 0

Lars Kotthoff
Lars Kotthoff

Reputation: 109262

This isn't implemented in mlr. We don't have any plans to support it, but you're always welcome to contribute a pull request.

Upvotes: 0

Related Questions