Similarity modifiers for enhancing the recommender system performance

Abstract

Matching users or items is the cornerstone for the success story of the recommender systems. This importance drives many efforts for enhancing the performance of the matching process in two directions. One direction investigates proposing new similarity measures, while the other one reshapes the existing ones using additional factors called modifiers to look at some statistical information of the two users into consideration. Sometimes, the modified versions are named new; however, they are just modified versions and can be used for other similarity measures as well. This paper studies current modifiers and proposes new ones. All are added to three well-known similarity measures and examined on three popular datasets. The results show that the weighing similarity modifiers enhance the performance of all the contested similarity measures. Two of the proposed modifiers show a significant improvement in prediction accuracy and error, especially for the users assisted by a few neighbors. They help the active users from the beginning, even with only one or two neighbors. Hence, this performance reflects the user preferences for items via the early election of trustworthy neighbors.