Collaborative Filtering with Implicit Feedbacks by Discounting Positive Feedbacks

Abstract

Recommender Systems are indispensable to provide personalized services on the Web. Recommending items which match a user's preference has been researched for a long time, and there exist a lot of useful approaches. Especially, Collaborative Filtering, which gives recommendation based on users' feedbacks to items, is considered useful. Feedbacks are categorized into explicit feedbacks and implicit feedbacks. In this paper, Collaborative Filtering with implicit feedbacks is addressed. Explicit feedbacks are feedbacks provided by users intentionally and represent users' preferences for items explicitly. For example, in Netflix, users can rate movies on a scale of 1-5, and, based on these ratings, users can receive movie recommendation. On the other hand, implicit feedbacks are collected by the system automatically. In Amazon.com, products that users buy and click are used for recommendation. While Collaborative Filtering with explicit feedbacks has been a central topic for a long time, implicit feedbacks have become a more and more important research topic recently because these are easier to obtain and more abundant than explicit feedbacks. However, implicit feedbacks are often noisy. They often contain feedbacks which do not represent users' real preferences for items. Our approach addresses to this noise problem. We propose three discounting methods for observed values in implicit feedbacks. The key idea is that there is hidden uncertainty for each observed feedback, and effects by observed feedbacks of much uncertainty are discounted. The three discounting methods do not need additional information besides ordinary user-item feedbacks pairs and timestamps. Experiments with huge real-world datasets confirm that all of the three methods contribute to improving the performance. Moreover, our discounting methods can easily be combined with existing methods and improve the recommendation accuracy of existing models.