Collaborative Filtering Using Projective Restoration Operators

We propose a modified version of our collaborative filtering method using restoration operators, which was proposed in [6]. Our previous method was designed so as to minimize expected squared error of predictions for user’s ratings, and we experimentally showed that, for users who have evaluated only small number of items, mean squared error of our method is smaller than that of correlation-base methods. After further experiments, however, we found that, for users who have evaluated many items, the best correlation-based method has smaller mean squared error than our method. In our modified version, we incorporated an idea of projecting on a low-dimensional subspace with our method using restoration operators. We experimentally showed that our modification overcame the shortcoming stated above.