[c++] Add Bagging by Query for Lambdarank

[shiyu1994]

Add bagging by query instead of by items in lambdarank, suggested by @metpavel. This should be more reasonable for bagging in ranking tasks. For a comparison of performance, on MS LTR dataset:
With bagging_freq=1 and bagging_fraction=0.1, if bagging_by_query=true

[LightGBM] [Info] Iteration:100, training ndcg@1 : 0.528525
[LightGBM] [Info] Iteration:100, training ndcg@3 : 0.502271
[LightGBM] [Info] Iteration:100, training ndcg@5 : 0.5034
[LightGBM] [Info] 23.123690 seconds elapsed, finished iteration 100
[LightGBM] [Info] Finished training

and if bagging_by_query=false

[LightGBM] [Info] Iteration:100, training ndcg@1 : 0.524889
[LightGBM] [Info] Iteration:100, training ndcg@3 : 0.502272
[LightGBM] [Info] Iteration:100, training ndcg@5 : 0.502838
[LightGBM] [Info] 43.811966 seconds elapsed, finished iteration 100
[LightGBM] [Info] Finished training

Without bagging

[LightGBM] [Info] Iteration:100, training ndcg@1 : 0.535041
[LightGBM] [Info] Iteration:100, training ndcg@3 : 0.509657
[LightGBM] [Info] Iteration:100, training ndcg@5 : 0.510785
[LightGBM] [Info] 50.232102 seconds elapsed, finished iteration 100
[LightGBM] [Info] Finished training

[borchero]

@neNasko1 is this something that might be missing for CUDA support in #6586?

[shiyu1994]

@guolinke Could you please help to review this when you have time? Thanks.

[StrikerRUS]

Just some very minor suggestions from me below:

[StrikerRUS]

PR's description states that bagging_by_query=True should improve metrics, but I don't see any comparison of bagging_by_query=True and bagging_by_query=False here...

[shiyu1994]

Since I found the result can be random when the dataset is small. For example, on CPU, bagging_by_query=True gets higher NDCG with the toy test dataset (even higher than the case without bagging), while with GPU bagging_by_query=True could get worse results compared with bagging_by_query=False. But when the dataset is large, for example, with MS LTR dataset, the results are less random, and bagging_by_query=True should improve performance, as in the description of this PR.

[shiyu1994]

In addition, we also see a significant improvement in training speed with bagging_by_query=True.

[StrikerRUS]

OK, I see. So this test is for something like "bagging_by_query=True doesn't break training".

[StrikerRUS]

LGTM, thanks!

[github-actions]

This pull request has been automatically locked since there has not been any recent activity since it was closed.
To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.