LTR with symmetric relevance on top and bottom elements of each query.

[jaguerrerod]

Is there a way to perform learning to rank so that documents at the beginning and end of each query are relevant in a symmetrical manner? The actual metric is Spearman correlation, and since LightGBM doesn't have a pairwise objective, I want to try an alternative by manipulating the relevances symmetrically with respect to the central part of each query.

[jameslamb]

@shiyu1994 could you please help with this question?

[jaguerrerod]

In the simplest scenario, every position carries equal weight, meaning that the crucial factor is the agreement in order between the responses and the predictions for every pair of observations. If the response Y is less than Z, then the prediction for Y (P(Y)) should also be less than the prediction for Z (P(Z)).
In other words, spearman rank correlation objective.