Can Latent Features Be Interpreted as Users in Matrix Factorization-Based Recommender Systems?

Matrix factorization has proven to be one of the most accurate recommendation approach. However, it faces one main shortcoming: the latent features that result from factorization, and that represent the underlying relation between users and items, are not directly interpretable. Some works focused on their interpretation, particularly with non-negative matrix factorization. In these works, features are viewed as groups of users, groups of items or as attributes of items, but such interpretations require human expertise. In this paper, we propose to interpret features as users, referred to as representative users. This interpretation relies on the study of the matrices that result from the factorization and on their link with the original rating matrix. Such an interpretation is not only performed automatically, as it does not require any human expertise, but it helps also to explain the recommendations made to users. In addition, we see it as a way to alleviate the new item cold-start problem, without requiring any information about the content of the items. The experiments conducted on several benchmark datasets confirm that the features discovered by a non-negative matrix factorization can be actually interpreted as users and that the representative users (the interpretations of the features), are a reliable source of information that allows to accurately estimate ratings on new items. They are thus a promising way to solve the new item cold-start problem.