Xgboost spark: Support XGBRanker

[WeichenXu123]

Xgboost spark: Support XGBRanker

[WeichenXu123]

@trivialfis @wbo4958 Do you have time to add this feature ? :)

[WeichenXu123]

@wbo4958 Do you have time to implement this feature ? @mengxr hopes this feature to go into xgboost 2.0.
Thanks!

[wbo4958]

Hi @WeichenXu123, I am pretty busy recently. I can't guarantee that I can make it in xgboost 2.0

[WeichenXu123]

@wbo4958 Understood. I can do it instead. Could you give the ETA of xgboost 2.0 releasing ?

[WeichenXu123]

@dbczumar Heads-up: I might need to spend some additional time to work on this because of above reason. :)

[trivialfis]

We are doing feature-based releases.

[WeichenXu123]

@trivialfis I have a question, in XGBRanker distributed training, shall we ensure each group training examples being put in the same worker data split ? And does training examples need to be sorted by "qid" first ?
(I will use "qid" not "group" param when constructing DMatrix)

[trivialfis]

Apologies for missing the thread.

And does training examples need to be sorted by "qid" first ?

yes.

shall we ensure each group training examples being put in the same worker data split

No

[wbo4958]

@WeichenXu123 @trivialfis Sorry, I didn't see #8172 has sorted the qid first? please correct me.

[trivialfis]

Users can have it sorted. Otherwise there will be an error in c++.

[silpara]

I don't see any documentation but SparkXGBRanker class is implemented at https://github.com/dmlc/xgboost/blob/master/python-package/xgboost/spark/estimator.py. Is it ready for use?

[trivialfis]

Opened a PR for the doc #8692 .

It's currently experimental.