Supervised Learning in Matrix Completion Framework for Recommender System Design

Recommender systems primarily utilize the, highly sparse, explicit rating information to make relevant predictions. This data scarcity places a limit on the accuracy of prediction. In this work we attempt to alleviate the problem of data sparsity by using secondary information. Most existing works incorporate auxiliary information in a (bi-linear) matrix factorization setup; whilst our model is based on a (convex) matrix completion framework. In this work, we use auxiliary information about users and items to impose additional constraints on the recovered rating values; adopting ideas from supervised learning. Alongside, we also propose a method to utilize the information map extracted from supervised learning approach to handle the cold start problem. Most works that address the cold start problem are focused on users with very few ratings this is not the pure cold-start problem. However, in this work we target new users and items which have no ratings available for them; and only has the associated metadata. We propose an algorithm using split Bregman technique for solving our formulations. Comparison of our design with existing state of the art methods for RS design on the movie recommender systems clearly indicate the superiority of our formulation over existing methods.