Making aggregate-level predictions in recommender systems using multi-level ratings

Aggregate-level ratings have been studied in recommender systems and have been shown to improve predictions of ratings of individual items for individual users. Similarly, individual-level ratings have also been used for the estimation of aggregate-level ratings for groups of items and users. In this paper, we combine these approaches and present a novel method for estimating unknown aggregate-level ratings from the known individual- And the aggregate-level ratings corresponding to different levels of the rating aggregation hierarchy. We show both theoretically and empirically that this combined approach outperforms the alternative methods that do not include the ratings from different levels of the rating aggregation hierarchy.