Question

I am implementing a user based reccomender that should work just on categories of items in order to avoid computation on useless data. To be more clear if a user is in a category page, I don’t want to calculate the preferences on all the items but just on users that have reviewed something in the same category.

To do this I am thinking to implement a custom UserSimilarity to tell Mahout that an user is not similar if he has never reviewed an item in this category. I can do the same implementing a Rescorer with filtered check.

I know that the userSimilarity is used at the beginning of the algorithm to avoid to make computation on the entire set but I don’t know if it is the same on the Rescorer. The Rescorer seems to me like a Post processor…I just want to avoid heavy computation. Should I use a Rescorer or a Custom UserSimilarity? How to choose between the two?

Answer 1

IDRescorer actually plays two roles, both as a way to pre-filter items to consider, and to post-process their scores. These correspond to its methods isFiltered() and rescore(). What you want to do is exactly what you do with isFiltered(). Return true for anything not in the category you are considering.