How does "label_gain" affect evaluation metrics: Relative or Absolute?

[unbreading]

Does the calculation of the NDCG metric depend on the absolute values of label_gain, or does it only consider their relative values?

For example, given labels 0, 1, and 2, I tried calculating NDCG with the following label_gain settings:

A. label_gain = [0, 1, 1]
B. label_gain = [0, 1, 1 + 1e-10]
C. label_gain = [0, 1, 2]

I observed that the difference in metrics between A and B is significant, while the difference between B and C is minimal or nonexistent. Does this imply that during evaluation, the NDCG metric is only influenced by the relative values of label_gain?

[unbreading]

Well, label_gain is exactly the 'gain' of ndcg. See:
https://github.com/microsoft/LightGBM/blob/master/src/metric/dcg_calculator.cpp#L109
https://github.com/microsoft/LightGBM/blob/master/src/metric/rank_metric.hpp#L86