Cold-start link prediction in multi-relational networks based on network dependence analysis

Cold-start link prediction has been a hot issue in complex network. Different with most of existing methods, this paper utilizes multiple interactions to predict a specific type of links. In this paper, multiple interactions are abstracted as multi-relational networks, and robust principle component analysis is employed to extract low-dimensional latent factors from sub-networks. Then a distribution free independence test, projection correlation, is introduced to efficiently analyze dependence between target and auxiliary sub-networks. Furthermore, associated auxiliary networks are exploited for cold-start link prediction, which aims to forecast potential links for new/isolated nodes in target sub-networks. Experimental results on 8 bioinformatics datasets validate rationality and effectiveness of the method.