Relevance meets calibration: Triple calibration distance design for neighbour-based recommender systems

Abstract

Calibrated recommendations are devoted to revealing the various preferences of users with the appropriate proportions in the recommendation list. Most of the existing calibrated-oriented recommendations take an extra postprocessing step to rerank the initial outputs. However, applying this postprocessing strategy may decrease the recommendation relevance, since the origin accurate outputs have been scattered, and they usually ignore the calibration between pairwise users/items. Instead of reranking the recommendation outputs, this article is dedicated to modifying the criterion of neighbour users’ selection, where we look forward to strengthening the recommendation relevance by calibrating the neighbourhood. We propose the first-order, second-order and the third-order calibration distance based on the motivation that if a user has a similar genre distribution or genre rating schema towards the target user, then his or her suggestions will be more useful for rating prediction. We also provide an equivalent transformation for the original method to speed up the algorithm with solid theoretical proof. Experimental analysis on two publicly available data sets empirically shows that our approaches are better than some of the state-of-the-art methods in terms of recommendation relevance, calibration and efficiency.