Improving Collaborative Filtering’s Rating Prediction Accuracy by Introducing the Common Item Rating Past Criterion

Collaborative filtering formulates personalized recommendations by considering ratings submitted by users. Collaborative filtering algorithms initially find people having similar likings, by inspecting the similarity of ratings already present in the ratings database. Users exhibiting high similarity regarding their likings are classified as “near neighbors” (NNs) and the ratings entered by each user's near neighbors drive the formulation of recommendations for that user. To quantify the similarity between users, in order to determine a user's NNs, a similarity metric is used. Insofar, similarity metrics proposed in the literature either consider all user ratings equally or take into account temporal variations within the users' or items' ratings history. However users' ratings are co-shaped according to the experiences that they had in the past; therefore if two users enter similar (or dissimilar) ratings for an item while having experienced to a large extent the same items in the past, this constitutes stronger evidence about user similarity (or dissimilarity). Insofar however, no similarity metric takes into account this aspect. In this work, we propose and evaluate an algorithm that considers the common item rating past when computing rating predictions, in order to increase rating prediction accuracy.