Back to archive

Thread

3 tweets

2
@paul_rietschka Yeah I think what everyone is aware of but nobody really knows how to deal with: - obviously you only get data for stuff that people see - strong position bias (especially when you are only showing a small number of recos) - over time it becomes a self-fulfilling prophecy.
3
@paul_rietschka What people do: - sprinkle in random items to get data. But this costs because those are almost always bad. - trying to account for position bias by reweighting - maybe quickly changing assortment actually helps, but what to do about cold start problems?