Having tried some of movie recommendation engines available on the web I have the feeling they are not satisfactory. I just fail to get movies similar to those I like based on traits interested for me personally.
My guess is that the lack of precision can be overcome with extending the number of latent factors in CF SVD algorithm. Another idea is that latent space can be initially hand-crafted.
E.g. we can take 50 well-studied movies and asses them by 200 traits manually (partially fill item-latent factors matrix). Accept this values as fixed. Then having some of the user ratings we can perform matrix factorization and inherit user-factors and utility matrices as usual, while keeping fixed values.
Any present work on this?