Ranking prediction are all same value

[1a1a11a]

I am using the XGBoost C API to do some ranking tasks and I have tried to use rank:pairwise, rank:map, and rank:ndcg as the objective, however, when I use rank:map, and rank:ndcg to train and predict, all the predicted values are constant, but it is not a problem when I use rank:pairwise. Does anyone know what could be wrong in my setup? Or is there a better way other than digging into the source code? Thank you!

[1a1a11a]

I have checked that the relevance in the training data are all from 0 to 1, and the inference data are not constant. I think the training has some problems, the metric does not change at all from the first iteration to 10 iterations (while it changes if I use
regression or rank:pairwise.

[creat89]

For NDCG, and I'm guessing MAP too, the relevance scores must be integers. See:

#6352 (comment)

or:

https://discuss.xgboost.ai/t/very-large-ndcg-result/1712