Using ndcg_exp_gain=False for non-exponential nDCG loss in XGBoost rank:ndcg

[HemakshiShardha]

I'm using XGBoost version 3.0.0 to train a Learning to Rank model with the rank:ndcg objective. I would like to train the model using the non-exponential (linear gain) version of the nDCG listwise loss, so I set the parameter:
ndcg_exp_gain = False

I’m posting to confirm a couple of things:

1. Is setting ndcg_exp_gain=False the only change required to switch from exponential to non-exponential (linear gain) nDCG loss during training?
2. Are there any known caveats, limitations, or gotchas associated with using ndcg_exp_gain=False that I should be aware of?

Any clarification or best practices would be greatly appreciated!

[0trm]

For the objective, yes. ndcg_exp_gain=False swaps the gain from 2^rel - 1 to rel in both the pairwise delta NDCG and the IDCG that normalizes it, and nothing else needs to change. It also drops the label checks, so relevance no longer has to be an integer of at most 31.

The catch on 3.0.0 is the eval metric. The built-in ndcg metric, which is also the default for rank:ndcg, didn't read that flag until 3.0.3 (#11534, fixing #11517). So on 3.0.0 you train on linear gain, but the ndcg@k you watch and early-stop on is still the exponential one, and with labels above 31 the metric fails outright as soon as you pass an eval set. I'd upgrade to 3.0.3 or later rather than work around it.

Two smaller things. With binary labels the flag changes nothing, since 2^1 - 1 = 1, so it only matters for graded labels. There, linear gain treats every grade step alike (a 4 ranked below a 3 costs the same as a 1 below a 0, at the same positions), while exponential gain puts most of the weight on getting the top grades first. I'd choose based on how much more a perfect result is worth to your users than a decent one. Also, with exp gain off the non-negative label check goes away too, so validate labels ≥ 0 yourself.

I read the 3.0.0 and 3.4.2 source for this but haven't run it.