Fusion of Similarity Measures in Recommender Systems

Abstract

Recommender systems provide suggestions for goods (such as movies or products) that a user may likely enjoy, taking into account their past preferences. An essential component of these systems is computing the similarity scores among users in order to detect individuals who share similar preferences. In order to enhance the precision of computing user similarities, this work introduces a novel “Fusion of Similarity Measures in Recommender Systems”. Contemporary similarity metrics such as Pearson correlation exhibit certain limitations. Sometimes, they exhibit a strong correlation between users despite significantly varied ratings, or a weak correlation for users with almost identical ratings. Furthermore, they fail to adequately consider variables such as the proportion of items that are co-rated by two users. Our approach computes user similarity in a novel manner that specifically tackles these concerns. It incorporates many parameters including rating closeness (how near the ratings are), rating significance (how distant from the average rating), rating singularity (the uniqueness of the ratings), and proportion of co-ratings. It also adapts for individual users' rating tendencies. Experiments demonstrate that our approach yields more precise similarity ratings then current approaches. It accurately classifies users as having high or low similarity depending on the measurable level of preference alignment. This significantly improves the quality of suggestions provided to users by the recommender system.