A comparative study of heterogeneous item recommendations in social systems

Abstract

While recommendation approaches exploiting different input sources have started to proliferate in the literature, an explicit study of the effect of the combination of heterogeneous inputs is still missing. On the other hand, in this context there are sides to recommendation quality requiring further characterisation and methodological research – a gap that is acknowledged in the field. We present a comparative study on the influence that different types of information available in social systems have on item recommendation. Aiming to identify which sources of user interest evidence – tags, social contacts, and user-item interaction data – are more effective to achieve useful recommendations, and in what aspect, we evaluate a number of content-based, collaborative filtering, and social recommenders on three datasets obtained from Delicious, Last.fm, and MovieLens. Aiming to determine whether and how combining such information sources may enhance over individual recommendation approaches, we extend the common accuracy-oriented evaluation practice with various metrics to measure further recommendation quality dimensions, namely coverage, diversity, novelty, overlap, and relative diversity between ranked item recommendations. We report empiric observations showing that exploiting tagging information by content-based recommenders provides high coverage and novelty, and combining social networking and collaborative filtering information by hybrid recommenders results in high accuracy and diversity. This, along with the fact that recommendation lists from the evaluated approaches had low overlap and relative diversity values between them, gives insights that meta-hybrid recommenders combining the above strategies may provide valuable, balanced item suggestions in terms of performance and non-performance metrics.