This paper considers a number of ways to introduce context into recommender systems. Specifically, it considers the factorization framework, in which a low-rank vector is computed for each user and each item. Preferences can be approximated by the scalar product of these vectors, although the paper also considers the cosine of the angle between vectors as a measure of preference because the scalar product method weighs more popular items more heavily.

Context as used in this paper is considered “as a property of a user-item interaction (e.g. time of the event, actual mood of the user when recommendation is requested, etc.).” This gives rise to a context feature matrix in addition to the item and user ones referred to earlier. The paper gives three formulas that model the predicted preferences of users for items. One of these ignores context, whereas the other two incorporate context in slightly different ways.

The authors ran experiments using four datasets (LastFM, TV1, TV2, and Grocery) to evaluate the effects of using these formulas. The evaluation of the results used recall and coverage, “a metric that quantifies the ability of the recommender system to explore the entire item catalog,” to evaluate the recommendations made using each formula. The results show that each approach has its strengths and weaknesses and that a clear winner cannot be decided. The authors discuss ways in which a user might choose between the methods. They indicate that predicting the most useful context dimension is a subject for future research.