[
https://issues.apache.org/jira/browse/SPARK-34415?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18097996#comment-18097996
]
Mao Li commented on SPARK-34415:
--------------------------------
Hi all — I'd like to pick this up, since it has been inactive since the 2021
revert
(SPARK-36664, PR #33819). [~phenry] please let me know if you still plan to
work on
this; otherwise I'll take it forward.
I've re-implemented ParamRandomBuilder from scratch, addressing the reasons for
the
revert directly:
1. True random sampling: build(numModels) draws each param independently for
every
candidate map (equivalent to sklearn's RandomizedSearchCV n_iter), instead
of the
original "sample n values per param, then take the Cartesian product".
2. Reproducibility contract ("MLlib algorithm's behavior should be fully
determined by
its params"): randomness is driven by an explicit constructor seed; same
seed + same
add* calls produce identical ParamMaps, verified by a dedicated test.
3. Narrower API surface: the Limits/Generator/RandomT type-class machinery is
not
revived; only concrete-typed addRandom/addLog10Random/addChoice overloads
are public
(which also keeps the API Java-friendly), with sampling internals
private[ml].
Plan: Scala-first PR (builder + tests), then follow-ups for PySpark parity, Java
example and docs — keeping the initial diff small and focused on the API
discussion.
Will open the PR shortly.
> Use randomization as a possibly better technique than grid search in
> optimizing hyperparameters
> -----------------------------------------------------------------------------------------------
>
> Key: SPARK-34415
> URL: https://issues.apache.org/jira/browse/SPARK-34415
> Project: Spark
> Issue Type: New Feature
> Components: ML, MLlib
> Affects Versions: 3.0.1
> Reporter: Phillip Henry
> Assignee: Phillip Henry
> Priority: Minor
>
> Randomization can be a more effective techinique than a grid search in
> finding optimal hyperparameters since min/max points can fall between the
> grid lines and never be found. Randomisation is not so restricted although
> the probability of finding minima/maxima is dependent on the number of
> attempts.
> Alice Zheng has an accessible description on how this technique works at
> [https://www.oreilly.com/library/view/evaluating-machine-learning/9781492048756/ch04.html]
> (Note that I have a PR for this work outstanding at
> [https://github.com/apache/spark/pull/31535] )
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
