← All questions
EvaluationVery Hard· Dwell Time· Regression· Learning to Rank

A colleague points out that your new labels are dwell t...

Switching to regression does not solve the loss-metric mismatch, as dwell time still reaches the metric through argsort. Furthermore, dwell time is not comparable across queries—30 seconds on a recipe is different from 30 seconds on a troubleshooting page—and it is heavily biased by the current ranker’s exposure policy. It is better used as a feature or for pairwise preference learning.

Code playground (Python)