0
$\begingroup$

I have a LightGBM Regressor model with 15 features. 5 of these features have 98.7% NA for the training set. All five of the features are NA for each row. I impute the missing values with zero before I feed into the model. The testing set has 90% NA for the same five features.

As a general question, how does this impact the decision making of the tree and the prediction from the model?

Since the testing set also is missing the majority of these 5 features, does it not have much of an effect on the model? It is my understanding that the model chooses the most important features for each prediction. Thus, I would assume that it uses the 5 features for instances that have them, and does not for the instances that do not.

$\endgroup$
1
  • $\begingroup$ What do the non-missing values in those columns look like? $\endgroup$
    – Ben Reiniger
    Commented Sep 6, 2023 at 13:04

1 Answer 1

1
$\begingroup$

All five of the features are NA for each row. I impute the missing values with zero before I feed into the model.

By setting them 0, the model will actually learn that the values are zero. I would consider you dropping these columns altogether.

$\endgroup$

Not the answer you're looking for? Browse other questions tagged or ask your own question.