Using past-prediction accuracy in recommender systems

This paper presents a new approach for memory-based collaborative filtering algorithms. In general, user-based rating prediction is a process in which each neighbor suggests a rating for the target item and the suggestions are combined by weighting the contribution of each neighbor. We present a new alternative that is independent of user rating scales and is based on what we call predictive probabilities. We explore how these probabilities can be used to select nearest neighbors for recommendations and integrate different types of dependence in the ratings. The neighborhood selection criterion depends on the capability of a user to predict past ratings. Our hypothesis is that if a user was good when predicting past ratings for an active user, then his predictions will also be helpful in recommending ratings for the same user in the future.