How to specify offset_column in h2o.stackedEnsemble() - r

Question

I am running gbm and glm with offset_column as base learners in h2o. My response variable is binary and the offset_column is a positive constant. Base learners worked. Here is the code:
train["offset"]<-train["log_hazard"] # offset column in the training set
my_gbm <- h2o.gbm(x = x, y = y, training_frame = train,
fold_column = "fold_id",
keep_cross_validation_predictions = TRUE,
offset_column = "offset",
seed = 1)
my_glm <- h2o.glm(x = x, y = y, training_frame = train,
fold_column = "fold_id",
keep_cross_validation_predictions = TRUE,
offset_column = "offset",
seed = 1,family = "binomial")
Then I am passing the offset_column in h2o.stackedEnsemble() through metalerner_params. Here is the code:
stack_model <- h2o.stackedEnsemble(x = x,
y = y,
training_frame = train,
base_models = list(my_gbm, my_glm),
metalearner_params = list(offset_column = "offset"))
But I received the following error:
ERRR on field: _offset_column: Offset column 'offset' not found in the training frame
The offset_column is in the training data. I am not sure why I am receiving this error message.
Then I tried running h2o.stackedEnsemble() without the metalerner_params option. Here is the code:
stack_model <- h2o.stackedEnsemble(x = x,
y = y,
training_frame = train,
base_models = list(my_gbm, my_glm))
and received the following warning message:
Warning message:
In .h2o.startModelJob(algo, params, h2oRestApiVersion) :
Dropping bad and constant columns: [offset].
I am not sure whether it ran properly. Can anyone please help me with this issue?

Answer 1

if you carefully read h2o docs for h2o.stackedEnsemble then you realize that h2o metalearner won't need offset parameter anymore as it will use cross-validated predicted values from base models to train:
my_gbm <- h2o.gbm(x = x, y = y, training_frame = train,
fold_column = "fold_id",
keep_cross_validation_predictions = TRUE,
offset_column = "offset",
seed = 1)
my_glm <- h2o.glm(x = x, y = y, training_frame = train,
fold_column = "fold_id",
keep_cross_validation_predictions = TRUE,
offset_column = "offset",
seed = 1,family = "binomial")
stack_model <- h2o.stackedEnsemble(x = x,
y = y,
training_frame = train,
base_models = list(my_gbm, my_glm))
h2o.performance(my_gbm, newdata = test)
h2o.performance(my_glm, newdata = test)
h2o.performance(stack_model, newdata = test)