Switched from xgboost 0.81 to 0.82 utilizing spark. I am running a transform on a trained XGBoostClassificationModel (BinaryClassification). testDF is transformed using a pipelineModel just before the XGBoostClassificationModel transform.
Previously in xgboost 0.81, I could run ‘xgboostModel.transform(testDF)’ on a non-persisted Dataset object and receive proper results.
Whether I persist testDF or not, I still receive an AUC ~ 0.9
Now within xgboost 0.82,
If I perist testDF, I receive an AUC ~ 0.9
If I do not persist testDF, I receive an AUC ~ 0.5 (my decile capture rate is flat)