Keywords: matrix factorization, triad dependency, mixed-effect model, collaborative prediction, shared parameters
Predicting unknown links in multiple layers of networks is critical to infer the structure of the network system and could be the primary interest in many network data applications. However, most existing link prediction methods are mainly designed for a single network and are not suitable for the multiple interconnected networks. The limitations of the existing methods include that an unrealistic independence assumption among links which neglects the neighborhood or high-order dependency among links, and are incapable of fully incorporating the cross-layer interaction in predicting links among different networks. In this paper, we propose a multilayer embedding method for unknown link prediction in interconnected networks. Specifically, our method incorporates the local dependency among inter-layer links and intra-layer links to fully capture the cross-layer interaction. In addition, we deploy a collaborative predicting scheme for between-layer links to utilize information from similar background or neighborhood. Computationally, we propose a scalable algorithm to handle large networks. In simulations, we show that the proposed method improves prediction accuracy compared to existing popular link prediction algorithms. We apply the proposed method for employment recommendation networks in China.