Question on predictions from ranking


I have a conceptual question on learning-to-rank, specifically about prediction. For training I input the groups to determine which documents should be ranked relatively to each other. But for prediction, I can’t input groups (at least not in the Python interface).

My understanding is that I would input the same type of data as with train and get back a ranking for each group. Instead, I input just the rows without group information and get back a score close to 0.5. What does that score then represent? A global rank? A score of how relevant that document is, and I can just group the predictions myself and then rank according to that score?


@twiecki At the time of prediction, you will get back a real-numbered output for each data row. You can use this output to compute predicted ordering within each query group. For each query group, you should sort the data rows by the outputs, in descending order.


Thanks, that’s very helpful!