Semi-supervised link prediction based on non-negative matrix factorization for temporal networks

Abstract

Temporal link prediction is a critical issue in the field of network analysis, which predicts the future links in temporal networks. In order to facilitate the performance of temporal link prediction approach, we should fuse the topological and temporal properties. Here we propose a novel semi-supervised non-negative matrix factorization method for temporal link prediction. Potential useful prior information is obtained from community which naturally expresses topological structure of networks. Moreover, we capture the temporal information of networks by graph communicability. We factorize the communicability matrix respect to the temporal network by setting the historic networks as graph regularization and priors as node pair constraints. Extensive experiments on both synthetic and real-world networks demonstrate that the proposed method can improve the accuracy of temporal link prediction. Especially, our method performs stably when the wrong prior rate is up to 30%.