A Non-Negative Matrix-Factorization-Based Network Embedding Approach for Hybrid Recommender Systems

Abstract

Currently, recommender systems are used in almost all online applications. A lot of algorithms have been proposed to predict user's preference including content filtering, collaboration filtering and hybird model. Traditional hybrid recommender system needs to combine the advantages of single element matrix factorization to reduce time consumption, however, existing single element based collaborative filtering methods need to further optimize node representation through attribute information. To better represent users and items accurately, network embedding method aiming at mining the low dimensional vector embedding of nodes in the user-item bipartite graph of the recommender system, becomes more and more active. In this paper, we propose an non-negative matrix factorization based network embedding approach for hybrid recommender systems, which learning representation for users and items on the basis of three weight edges, and an objective function for network embedding, which is suitable for recommendation system. The function is obtained by jointly optimizing NMF approach, which enables the learned representations of nodes to preserve the microscopic of networks. In addition, we provide complete updating rules to compute parameters of our model, which adopts single element based non-negative matrix factorization model, and it is very suitable for recommender system integrating content-based and collaborative filtering, and improve the accuracy of the model and ensure that it finds the best point. Experimental results on a variety of real-world networks indicate that the performance of the proposed method is superior to the state-of-the-arts.