Enhancing recommender systems prediction through qualitative preference relations

In this work, we propose a novel approach to deal with user preference relations instead of absolute ratings, in recommender systems. User's preferences are ratings expressed qualitatively by using linguistic terms. This is a suitable technique when preferences are imprecise and vague. Due to the fact that the overall item rating may hide the users' preferences heterogeneity and mislead the system when predicting the items (products / services) that users are interested in, we also choose to incorporate multi-criteria ratings, which is a promising technique to improve the recommender systems accuracy. User's items ratings are represented through a preference graph which highlight better items relationships. Similarity between users is performed on the basis of the similarity of their preference relations instead of their absolute ratings, since preference relations can better reflect similar users' ratings patterns. Our approach enhances somehow the classical recommender system precision because the graphs used for prediction are more informative and reflect user's initial ratings relations.