Encoding temporal and structural information in machine learning models for recommendation

Recommender systems focus on the task of selecting relevant items for a user, according to their taste: movies to watch, books to read, etc. In this paper, we provide a proof-of-concept that graphs and link streams can be relevant for the recommendation task. We design simple features modelling the structure and its evolution with time of recommender systems in order to validate this claim. We learn these features alongside classical content-based features, using a gradient boosting machine model (XgBoost) and perform rating prediction on a movie rating dataset, MovieLens20M, and a book rating dataset, from GoodReads. We obtain comparable performance to the state-of-the-art, and some interesting leads in terms of explicability.