A New Model of Selecting Most Relevant Ratings in Recommender Systems

Abstract. A major assumption of collab-orative filtering is that similar users will always agree on a majority of items, re-gardless of their domain. This concept es-tablishes strong connections among neigh-bors. However, it eliminates potentially good users on the premise that they are not similar enough. Furthermore, this assump-tion allows for the possibility of a neighbor to be chosen simply because he/she shares a lot of similar ratings in unrelated domains and offers little useful information in the ac-tive item domain. This effectively reduces the amount of useful information that is considered for each recommendation. We propose a new way to identify relevant rat-ings that relies on somewhat weaker, but more abundantly available neighbors