A New Graph-based Approach to Recommender Systems

Abstract

Recommender Systems are responsible for providing recommendations to the users. They are extensively used in online marketplaces like Amazon.com, Netflix, etc. to increase revenue by enhancing the user experience, as they provide personalized lists of items. Recommendation accuracy is one of the widely accepted parameter to judge performance of recommender systems. Some of the conventional techniques to realize recommender systems include collaborative filtering and content based filtering. Graphbased recommender system have been analysed recently, to increase recommendation accuracy. In this research, we propose a matrix factorization based graph model that outperforms collaborative filtering and content-based filtering in terms of recommendation accuracy. Conventional recommender system approaches, however, do not focus on the consensus between users and items. Such consensus model makes sense in fields like job recommendations, where a candidate should be matched with a job which is satisfactory for both candidate and the employer. We propose a compatibility metric that models such scenario as a recommendation problem. Further, we propose a two-way rating graph model that addresses this issue. It finds a matching that increases compatibility value of the recommendations. We tested this proposed model on dating dataset made available by Occams Lab. It outperforms random generation of the recommendations in terms of compatibility value