I'm doing an experiment creating a recommender from a Pinterest crawl I have going. I have at least three actions that relate to recommendations:
Goal: recommend people you (a pinterest user) might want to follow Actions mined by crawling: follows (user, user) followed by (user, user) repinned (user, user) here a user repinned something from another user training a recommender on one of the three actions and measuring the precision of recommendations leads to different measures of predictive power for each of the three actions. So holding out follows each of the three actions predicts the held out follows with different precision. Since all three actions (an a few others I can think of) have user IDs of the primary user in common, they exist in the same user space and I can create a cross recommender ensemble (see previous cross recommender emails) so R_f + aR_fb + bR_rp = R. Now it is a matter of learning the weights a and b by finding the ones that create the highest ensemble precision score. Gradient assent anyone? Seem right? If I anonymize the already public data by hashing the usernames I might be able to share the data. Seems like I should be able to but if anyone knows one way or the other, please speak up.
