A novel hybrid approach improving effectiveness of recommender systems

摘要

Recommendation systems are an important part of many online services that we have grown accustomed to. They automatically suggest movies you may like to watch on Netflix, or products that you may otherwise not have noticed on Amazon. They are, however, very much still works in progress. Despite recent advances in recommendation systems, many fundamental, challenging problems still remain, such as the cold-boot problem (making good recommendations when there is a lack of sufficient data). This paper gives a good run-down of the many different types of recommendation systems available today, highlighting their strengths and weaknesses. Considering the state of the art, the author then proposes a hybrid system that taps on the collective benefits of content-based recommendation systems and collaborative filtering systems. Key to the proposal is an integration of "individual relevance" with "social relevance." The former captures how relevant an item is to a particular user, while the latter models the relevance of an item, taking into consideration the preferences of other like-minded users. Experiments that were run over two custom datasets showed that the proposed hybrid approach outperforms several baselines.