Basically, I am trying to have a recommender system based on SVD for a Boolean utility matrix. ie If at all some entries are present in the utility matrix, they will be 1 (I made it pseudo-implicit from explicit ratings available from the Movielens dataset). I used the 'irlba' package in R to compute the partial SVD quickly. My confusion is, whether the SVD is a good technique for boolean user-item preference matrix. I read online that SVD doesn't offer that many 'interpretable' latent factors/features for a Boolean sparse utility matrix. So, will it lead to some kind of an over fitting if the rank-k (the dimension of the latent factor space) is chosen as some value beyond say 20? I used the 10 Million ratings Movielens Data and modified them as 5 Million sparse 1 entries (for ratings= 4 or 5). The problem is, the top-N list from this model for many users do not match at all (at the maximum of 1 or 2 intersections) with the collaborative-user-neighborhood similarity based recommender system. Could someone please tell whether SVD can be applied to my situation ( I need to make the recommendations on boolean sparse utility matrix) and if yes, how should I modify my present approach?
Add a comment
|