Towards a Novel Graph-based collaborative filtering approach for recommendation systems

Abstract

Recommendation systems have become an entrenched part of emarket platforms as they help offer personalized user experience, increasing thus loyalty, satisfaction and lifetime value. Although great effort has been devoted to the proposal, implementation and study of recommendation systems approaches, there is still a lot of room for improvement. This work proposes a novel Graph-based Collaborative Filtering approach for recommendation systems. The key idea is to improve the accuracy of Hybrid approaches by basing recommendations on a more refined Homophily (similarity) level. We hypothesize that a user-base is a network of highly similar sub-communities and that any inference should be made on these communities' level rather than on the whole user-base. Furthermore, some users are more embedded in these similarity communities than others. We think that these "Key-nodes" can lead to more accurate recommendations than the ones based on aggregations on similar users. The paper also presents a validation of our approach and the computation of the accuracy of its underlying model over the classical MovieLens dataset. The experiment's results indicate significant gain in performance compared to two classical recommendation systems' approaches.