How to set qid when predict in LGBMRanker?

[deepfuzzy]

In python, qid can be set to group in fit(X, y, group), but predict() has no parameter group, so where to set qid?

[guolinke]

qid is not needed in prediction. LightGBM will predict the relevant score, and you should rank the result by yourself.

[mlisovyi]

@guolinke is it possible to elaborate more on this? Shall one make predictions on each group separately? What is the output of the LGBM.predict() method? The docs are very short on the subject and examples in the tests and elsewhere address rnker fit, but not prediction

[guolinke]

@mlisovyi qid have no effects on ranking predictions in lgb.predict, as the predicted results are the relevant scores. And you can use the relevant scores to get the correct ranking order results by yourself, and qid is needed in this step.

[mlisovyi]

I see... How to treat negative values in the predictions?

[guolinke]

@mlisovyi the same, you can just sort by prediction values.

[nonva]

but, not in same query group, is the scores can be comparable?

[guolinke]

@nonva theoretically, they could be comparable.