I’m tuning a binary model with PROC GRADBOOST + AUTOTUNE and PARTITION (train/validation/test). The HISTORYTABLE= captures one objective per evaluation (Gini on validation or the k-fold average), but I need Gini per partition (train/valid/test) for every evaluated configuration. Right now, my workaround is to refit each configuration, score, and run PROC ASSESS BY _PartInd_ to compute AUC → Gini, which is costly.
Questions:
- Is there a built-in way to get partition-level metrics (train/valid/test) for all evaluated configurations without refitting each one?
- Is there a supported pattern to persist each candidate model during tuning (not just the champion) so I can score later and compute partition metrics without retraining?
- Any roadmap to add partition-level metrics (or an option to save all intermediate models) to AUTOTUNE outputs?
From the docs, AUTOTUNE records a single objective per evaluation (validation/k-fold) in the history and doesn’t expose per-partition metrics by default.
