Is it ok to retweak item_similarity_recommender.create to work on user-user similarity instead?

User 3230 | 2/20/2016, 9:38:26 AM

In my dataset, it is inherently more likely that a user-user recommendation system will provide better precision and recall.

I could "rotate" the user-rating matrix by swapping the userid and itemid columns both in the train/test split command and the eventual create command.

I could then use getsimilarusers to retrieve the actual similar items, would this work? Or is there a better way to create a user-user recommendation model in Graphlabs?


User 1592 | 2/20/2016, 10:52:41 AM

Hi Absolutely, this is the way to go. Just swap the userid and itemid columns and it will work.