179 questions
No questions match those filters.
← All questions
EvaluationHard· NDCG· Learning to Rank· Loss Functions
You have query–document pairs graded 0 to 4 by humans,...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansThe core problem is that nDCG is a function of the relative ordering (argsort) of the scores rather than their absolute values. Because of this, a model can achieve a perfect nDCG of 1.000 while still maintaining a high mean squared error. Consequently, the training loss can decrease significantly without producing any meaningful improvement in the actual ranking quality, as the model is optimizing for absolute values that do not impact the final metric.