A fusion probability matrix factorization framework for link prediction

Link prediction is a fundamental research problem in network data analysis. Networks usually contain rich node to-node topological metrics and their effective use is crucial to solve the link prediction problem. Despite significant advances, the existing metric-based link prediction methods usually only consider one single topological metric and thus show some limitations in different types of networks; the existing matrix factorization-based models mainly focus on modeling the adjacent matrix of a network, and this is hard to ensure the modeling of those topological metrics that can play an important role in link prediction. This study develops effective approaches by fusing the adjacent matrix and some key topological metrics in a unified probability matrix factorization framework. In these approaches, we consider not only the symmetric metrics but also the asymmetric metrics which are usually not taken into consideration in the related work. In our probability matrix factorization framework, we first present two fusion models by fusing two kinds of metrics respectively, and based on the fusion models, we put forward the final fusion models which fuse the two kinds of metrics simultaneously. To verify the performance of all the fusion models, we conduct the experiments with six directed networks and six undirected ones, and the extensive experiments show that the proposed models provide impressive predicting performance for link prediction.