Practically Feasible Recommender Systems for Cold Start Problems

Recommender systems are beneficial to both service providers and users, as they offer item recommendations to individual users based on their preferences. However, the recommender systems present cold start problems, which refer to the fact that it is difficult to recommend new items for newly added users. Recommender systems for existing users and items have been researched vigorously. On the other hand, cold start problems have received insufficient attention. In this paper, we propose methods for recommender filtering that can be applied to address cold start problems. Our models provide simple matrix approximation methods for analyzing large volumes of unlabeled data from potential items. We use matrix factorization techniques based on a pseudo-inverse matrix for rating approximation, which enables us to make cold start recommendations by using a user-characteristic matrix, item-feature matrix, and rating matrix. The proposed models can explicitly learn user demographics and item features to overcome the cold start problems. The results of a numerical experiment show that the proposed methods outperform several baselines when tested on the MovieLens 1M dataset.