Release highlights for 1.3

[jeremiedbb]

As usual, let's start with a few highlights and add more if needed in subsequent PRs.
For now I put the TargetEncoder, HDBSCAN, and missing values support in trees.

Do not hesitate to edit.

Something we should add is some news about the metadata routing but I wasn't sure how to write that. @adrinjalali would you mind adding a section (here or in a separate PR) ?

[jeremiedbb]

Other features that I'm thinking about are

• Grouping infrequent categories in OrdinalEncoder
• Gamma deviance in HGBRegressor

Do you think they should end up in the highlights ?

[adrinjalali]

@jeremiedbb added metadata routing.

[thomasjpfan]

I think this is a good starting point to iterate from. LGTM

[thomasjpfan]

Nit:

Suggested change

	v_measure_score(true_labels[hdbscan.labels_ != -1], non_noisy_labels)
	print(
	"V-measure:", v_measure_score(true_labels[hdbscan.labels_ != -1], non_noisy_labels)
	)

[lorentzenchr]

I would place this at the top.

[lorentzenchr]

Other features that I'm thinking about are

• Grouping infrequent categories in OrdinalEncoder
• Gamma deviance in HGBRegressor

From a practitioner‘s perspective, pricing actuaries in particular, gamma deviance HGBT are a big deal. The PR itself was pretty small, but based on a pile of work with the common loss functions.

[jeremiedbb]

@lorentzenchr would you mind adding a small section for the Gamma deviance ?

[github-actions]

✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: a7b9264. Link to the linter CI: here}

[Micky774]

The wording is a bit difficult here only because HDBSCAN doesn't solve exactly the DBSCAN problem for multiple epsilon, but rather a modified problem that solves for the same solutions locally. I think this is a fair-enough compromise between correctness and readability. Open to thoughts though :)

Suggested change

	# By performing :class:`cluster.DBSCAN` over varying epsilon values
	# :class:`cluster.HDBSCAN` finds clusters of varying densities making it more robust to
	# parameter selection than :class:`cluster.DBSCAN`. More details in the
	# :ref:`User Guide <hdbscan>`.
	# By performing a modified version of :class:`cluster.DBSCAN` over multiple epsilon
	# values simultaneously, :class:`cluster.HDBSCAN` finds clusters of varying densities
	# making it more robust to parameter selection than :class:`cluster.DBSCAN`.
	# More details in the :ref:`User Guide <hdbscan>`.

[lorentzenchr]

@lorentzenchr would you mind adding a small section for the Gamma deviance ?

Until when?
Something like https://scikit-learn.org/stable/auto_examples/release_highlights/plot_release_highlights_0_23_0.html#generalized-linear-models-and-poisson-loss-for-gradient-boosting will do with the rng.poisson replaced by rng.gamma.

[ogrisel]

LGTM!

[glemaitre]

LGTM on my side.

[lorentzenchr]

@jeremiedbb Thank you! 🚀