Exploring Explanations for Matrix Factorization Recommender Systems

Como, Italy

In this paper we address the problem of finding explanations for collaborative filtering algorithms that use matrix factorization methods. We look for explanations that increase the transparency of the system. To do so, we propose two measures. First, we show a model that describes the contribution of each previous rating given by a user to the generated recommendation. Second, we measure the influence of changing each previous rating of a user on the outcome of the recommender system. We show that under the assumption that there are many more users in the system than there are items, we can efficiently generate each type of explanation by using linear approximations of the recommender system’s behavior for each user, and computing partial derivatives of predicted ratings with respect to each user’s provided ratings.