Novel recommendation algorithm based on clustering dissimilarity measures

Abstract

With the rapid development of information technology and Internet computing, recommendation systems of various forms have been used in nearly all large-scale e-commerce platforms to different degrees. Besides accuracy, novelty also influences the satisfaction of users with these recommendation systems. To be considered novel, an item should show some differences from the preferences of users. To recommend diverse items under the precondition of guaranteeing accuracy, a clustering-based novel recommendation algorithm was proposed in this study. User preference was modeled by using global clustering and related-items clustering (RC), weighted distance was employed to calculate the dissimilarity of the to-be-recommended items from user preference, and item novelty was measured by combining satisfactory and unknown, followed by recommendation. The effectiveness of the proposed algorithm was verified by conducting an offline experiment. Results show that the recall rate of unknown items of RC increases by 15% whereas that of known items decreases by 60%, thereby indicating that the novelty of the recommendation result is remarkably improved. Meanwhile, the coverage rate increases by approximately 200%, thereby indicating that the proposed recommendation system demonstrates an improved ability to recommend long-tail items. This study provides references for improving the satisfaction of users with recommendation systems.