I have a dataset with a long-long tail, which means that I should get many cold-start problems after splitting the dataset.
I expected to have a low AUC, but I am getting a relatively high AUC (83%) on the test set.
So, how does surprise specifically deals with the cold-start problem? Does it recommend any (or random) item to a new user? Or, the new user is kept out of the metric calculation?
The algorithms in surprise don't have a fallback to specifically handle cold start but some (e.g. SVD) will naturally fallback to the mean of the ratings if there are no user-wise or item-wise info available
Hi Nicolas,
Surprise is great!
I have a dataset with a long-long tail, which means that I should get many cold-start problems after splitting the dataset.
I expected to have a low AUC, but I am getting a relatively high AUC (83%) on the test set.
So, how does surprise specifically deals with the cold-start problem? Does it recommend any (or random) item to a new user? Or, the new user is kept out of the metric calculation?
I couldn't find such info in the documents.
Thank you very much for your attention.
Regards,
Vitor