Reproducibility audit: remaining estimators (follow-up to #1364, #1369, #1374, #1376)

[immu4989]

Is your feature request related to a problem? Please describe.

No response

Describe the solution you'd like

Background

@dannycg1996 has merged a series of reproducibility fixes for FLAML estimators:

• fix: FLAML catboost metrics arent reproducible #1364 — CatBoost metrics reproducibility
• Flaml: fix lgbm reproducibility #1369 — LGBM reproducibility
• fix: Ensure Reproducibility of ElasticNetEstimator #1374 — ElasticNetEstimator reproducibility
• fix: LinearSVC results now reproducible #1376 — LinearSVC reproducibility

These follow up on the long-standing reproducibility concern raised in #151.

Proposal

I'd like to extend that work and audit the remaining estimators in
flaml/automl/model.py that accept a random_state. After reading the
merged PRs above to confirm the pattern, my plan is to submit one focused
PR per estimator (or per class family where inheritance makes a single PR
appropriate), each with:

• A reproducibility unit test in the style of the tests added in Flaml: fix lgbm reproducibility #1369 / fix: Ensure Reproducibility of ElasticNetEstimator #1374
• The minimal config / __init__ / config2params change required to
  ensure a seeded run produces identical results across repeated fits

Estimators not yet covered

Estimator	Notes
RandomForestEstimator	Has fixed default random_state since v1.1.0;
	want to verify end-to-end reproducibility
ExtraTreesEstimator	Inherits from RandomForestEstimator
XGBoostEstimator	Non-sklearn API path
XGBoostSklearnEstimator	sklearn API path
XGBoostLimitDepthEstimator	Inherits from XGBoostSklearnEstimator
LRL1Classifier	LR with L1 — relevant for liblinear/saga
LRL2Classifier	LR with L2 — relevant for sag/saga
SGDEstimator	Stochastic gradient descent

Questions for maintainers

1. Is this audit welcome, and is the per-estimator-PR approach preferred
   over a single combined PR?
2. Should the audit also cover the corresponding "flamlized" zero-shot
   estimators in flaml/default/estimator.py, or are those out of scope?
3. Are there estimators on the list above that you'd like me to skip
   (e.g., known-deprecated paths)?

I'd plan to start with RandomForestEstimator once you confirm the
direction. Happy to adjust scope based on your guidance.

Additional context

##Why this matters to me

In production ML for HR risk modeling, reproducibility is a regulatory
and audit requirement — being able to re-derive the same model from the
same data and config is non-negotiable. FLAML is increasingly used in
those settings (including via Microsoft Fabric), so closing this gap
across the estimator family seems worthwhile.

[thinkall]

Hi @immu4989 ,thank you very much for the proposal. PRs are super welcome. I think one PR for each estimator is better than putting everything in a single PR.

[immu4989]

Sure @thinkall . Thank you for the information. Submitted a pull request for SGD Estimator.