XGBRanker support

[klvnptr]

Hey Guys.

Just a quick question. Are you guys planning to support XGBRanker?
https://xgboost.readthedocs.io/en/stable/python/python_api.html#xgboost.XGBRanker

Thanks. Peter

[montanalow]

I'm guessing you would like listwise ranking objective, rather than the pointwise objective which basically degrades to ordering by regression. listwise ranking needs a batch prediction interface, since it takes an entire set of candidates into consideration, and it needs the additional query set size argument. We'll need to add a "ranking" objective in addition to classification and regression, for projects to support the additional API and data requirements. This seems like a pretty reasonable thing to support.

Do you have some benchmarks for the differences in quality you see between pointwise/pairwise/listwise ranking on your dataset so we have more context around the metrics you're looking at for this problem?

[klvnptr]

Thank you for your feedback. No real-world experience yet, we are just getting into this space. We are implementing pairwise right now.