Hi,
faced with interesting question, answer for which I don't have. When training LGBM Ranker it's necessary to provide a group array, which helps the model to distinct sessions and compare only those elements, which belong to the same group. So while training the model doesn't know that there are other groups when it compares documents from a given one. But while prediction we do not provide the model with group array and scores are not depended on the group. That's a little bit confusing. My question is - are the ranker scores for different groups in the same space or in other words - are scores for different groups comparable?