Abstract
In the current information overload context caused by the large volume of accessible digital data, recommender systems allow to guide the user in his/her learning, shopping, leisure, music listening, reading activities..., by suggesting personalized items. To do so, recommendation models predict users' preferences for their unrated items. Classical recommendation approaches, such as collaborative filtering, for example, rely on data collected through user feedback, usually in the form of a rating matrix, and try to discover relevant information to characterize and predict user tastes. In addition to the user feedback data, the knowledge related to items themselves also represents a major asset for improving the performance of recommendation systems. Knowledge engineering, more specifically the semantic Web and knowledge graphs, can play a central role. In this context, our research works propose different ways to improve recommendation systems, adopting a “from-data-to-knowledge” transversal vision, and consider three different recommendation aspects: accuracy, diversification and explicability.Our first contribution is mainly focused on pure user feedback data. It aims at improving the accuracy of recommendations in terms of the prediction of users’ tastes. We propose EBCR (Empirical Bayes Concordance Ratio), a simple and generic method inspired by Bayesian inference, which allows to adjust the similarity computations between users (or between items) in collaborative filtering, according to the number of co-rated items (or the number of users having rated the same item). Experiments conducted on benchmark datasets have confirmed that this method systematically improves the predictive accuracy of collaborative filtering for all considered similarity measures.Our second contribution concerns the diversification of recommendations. We have conducted an in-depth study to compare and analyze the performance of seven recommendation models including classical models such as collaborative filtering and latent factor models as well as more recent ones based on deep neural networks and knowledge graph embeddings. We have evaluated their ability to provide diversified items and proposed an approach that allows adjusting diversity to specific user needs. In order to estimate the diversity of recommendations, we considered semantic similarity measures by leveraging the semantic Web and knowledge graphs.Finally, our third contribution concerns the explicability of recommendations. Here, we further exploit domain knowledge and propose a post-hoc recommendation explanation approach that effectively accounts for the hierarchy of item properties within the DBpedia knowledge graph. Evaluation results of our approach based on an online user study including 155 participants suggest significant improvements in terms of engagement, trust and persuasion.