Improving the Performance of Collaborative Filtering Recommender Systems through User Profile Clustering

Recommender systems offer personalization to online activities due to their ability to recommend products that are unknown to the user. The most common form of these systems employs collaborative filtering to make recommendations and operate by estimating a preference for an item based on how like minded users have previously rated items. Such methods require large amounts of training data which highlights a scalability problem of collaborative filtering, namely, the trade-off between accurate estimation prediction and the time required to calculate them. This paper demonstrates a novel approach to determine interest thus improving scalability by partitioning training data into user based profile clusters. The partitioned data represents user segments (or profile types) which is used to as a more concise representation of similar users for the target. Experimental results have shown a dramatic increase in prediction speed without a loss in accuracy.