EBM pmml does not support missing values

[sadsquirrel369]

Hi there,

I came across a discrepancy between the pmml prediction for the ebm in a dataset where there are nulls. The ebm object has a score for nan's but the pmml file does not. Is there a way to fix this. This also becomes a problem where you have two variables interacting with Nan's.

Thank you

[vruusmann]

EBM is an ensemble model.

PMML uses the Segmentation@missingPredictionTreatment attribute to specify what to do in case some member model(s) of an ensemble model return a missing (sub-)prediction.

The current behaviour is returnMissing, which means "abandon the scoring process, and return a missing value as the final prediction".

You may change this attribute to skipSegment, which means "continue the scoring process, and ignore this missing sub-prediction when computing the final prediction".

Does this attribute change fix your issue, meaning that the EBM PMML starts making correct predictions? If not, then it must mean that EBM is performing some model-internal imputation. For example, maybe it replaces all NaN values with 0 values while ingesting features.

[vruusmann]

@sadsquirrel369 If you want to see this issue fixed soon, then it's your job to dig through the EBM codebase and identify what happens to those inputted NaN values. Paste your research results here.