Aggregation trade offs in family based recommendations

Abstract

Personalized information access tools are frequently based on collaborative filtering recommendation algorithms. Collaborative filtering recommender systems typically suffer from a data sparsity problem, where systems do not have sufficient user data to generate accurate and reliable predictions. Prior research suggested using group-based user data in the collaborative filtering recommendation process to generate group-based predictions and partially resolve the sparsity problem. Although group recommendations are less accurate than personalized recommendations, they are more accurate than general non-personalized recommendations, which are the natural fall back when personalized recommendations cannot be generated. In this work we present initial results of a study that exploits the browsing logs of real families of users gathered in an eHealth portal. The browsing logs allowed us to experimentally compare the accuracy of two group-based recommendation strategies: aggregated group models and aggregated predictions. Our results showed that aggregating individual models into group models resulted in more accurate predictions than aggregating individual predictions into group predictions. © Springer-Verlag Berlin Heidelberg 2009.