An Assessment of Collaborative Filtering-Based Recommender Systems : And their Ability to take the Individual Preferences of Users into Account

Abstract: This study investigated the effect that aggregation functions and similarity measures had on the accuracy of memory based collaborative filtering algorithms, specifically in the case when the actual rating that a user would provide deviates significantly from the rating that the average user would provide. Nine different combinations of algorithms were implemented by using three different aggregation functions and three different similarity measures. These algorithms were tested on the MovieLens dataset and the similarity measures used were the Pearson correlation, JPSS and JEPSS which is a modified version of JPSS. From the results it was concluded that using the normsum aggregation function lead to a major improvement in accuracy in general. Another conclusion was that using JEPSS over JPSS does not provide a significant increase in accuracy. Finally, the best results for JPSS and JEPSS were obtained when using the least amount of similar users.