Release v0.17.0

[jeongyoonlee]

Release v0.17.0

Bumps the version to 0.17.0 and adds the changelog entry. This is the release PR — once approved and merged, the v0.17.0 GitHub Release will be created, which triggers the PyPI publish workflow (cibuildwheel + trusted publishing).

Why now

The released 0.16.0 wheel crashes on import under scikit-learn 1.9 (ImportError: cannot import name 'DOUBLE' from 'sklearn.ensemble._forest', #926). The fix is already on master (#903 / #907 / #909) but hasn't shipped — cutting this release resolves #926 for all PyPI users. Verified on master + scikit-learn 1.9.0: import causalml.dataset and a CausalRandomForestRegressor fit/predict run clean.

Highlights (29 PRs since v0.16.0)

• scikit-learn 1.9 support (Fix for compatibility with sklearn v1.9.0 #903, Fix CausalRandomForestRegressor.fit() on scikit-learn 1.9 #907, Support CausalRandomForestRegressor.calculate_error() on scikit-learn >= 1.9 #909)
• Native Polars support across all meta-learners (Add native Polars DataFrame, Series, and LazyFrame support for all meta-learners #901, Polish Polars support and add bootstrap CI test coverage #921)
• JAX/flax.nnx backend for DragonNet (Add JAX/flax.nnx backend for DragonNet #918, docs: document JAX backend for DragonNet #919)
• scikit-learn-compliant meta-learners via BaseEstimator — clone()/get_params() work (Make meta-learners scikit-learn compliant via BaseEstimator #912)
• RATE metric + rate_score() with bootstrap CIs (Add Rank-weighted Average Treatment Effect (RATE) metric #887, Add rate_score() with bootstrap confidence intervals and p-values #890); post-fit CIs for BaseTLearner (Add post-fit confidence intervals to BaseTLearner via store_bootstraps and return_ci #886)
• Numerous bug fixes (see docs/changelog.rst)

⚠️ Breaking change

• Make meta-learners scikit-learn compliant via BaseEstimator #912: meta-learner __init__ now stores constructor args verbatim and builds models in fit(). Most visibly, XGBRRegressor no longer accepts arbitrary **kwargs — pass XGBoost params via the explicit xgb_kwargs=<dict> argument.

Full notes in docs/changelog.rst. Resolves #926 (on release).

[t-tte]

LGTM