Partitioning in Recommender Systems
Date of Submission:
April 2, 1998
Recommender systems help us overcome information overload by computing personalized recommendations for use of items we may find interesting. We consider two fundamental problems in recommender systems: how to improve the quality of the ratings, and how to improve the efficiency and scalability of the recommender system itself. Partitioning is a likely solution to both of these problems, becasue it reduces the size of the datasets, and because it may group related items together. We partition the items and users in various ways-randomly, by item content, and by a clustering algorithm. We find that partitioning items by content and by clusterng improves the quality of the recommendations, with clustering out-performing content. Partitioning users leads to improved efficiency and scalability, but at the cost of a slight loss of quality.