How do you determine the right number of item candidates to generate in the first stage? Any rules of thumb?
The competition metric is scored at 100 candidates, so its the minimun number to have. Increasing the number of candidates increases the recall score, but at the same time increases the dataset for the second stage (reranker). So we iterativelly experimented increased number of candidates checking accuracy and at the same time avoiding running out-of-memory at the second stage.
1 Like
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.