SemanticSVD++: Incorporating Semantic Taste Evolution for Predicting Ratings

Recommender systems profile the preferences of users and then use this information to forecast users´ future ratings. One of the most common recommendation approaches is the use of matrix factorisation in which users´ past ratings of items (i.e. Movies, books, etc.) are used to capture their affinity to implicit factors. A central limitation of such factorisation is that one cannot consider how a user´s preferences for a factor have changed over time. In this paper we present the SemanticSVD++ model that overcomes this limitation by using the semantic categories of recommendation items as prior factors for a given user. We present a model to capture the semantic taste evolution of users over time, and demonstrate how such development is susceptible to global influence dynamics. We explain how the SemanticSVD++ model incorporates such evolution information within a matrix factorisation approach, and empirically demonstrate the improvement in predictive capability that this yields when tested on two independent movie recommendation datasets.