A multilayered approach for link prediction in heterogeneous complex networks

Link prediction problem is a difficult task in complex networks due to (i) network size and sparsity, and (ii) extracting efficient similarity measures between node pairs. Although many link predictors have been proposed in the literatures, most of them are presented in homogeneous networks. The heterogeneity of node types adds a new challenge to link prediction problem. A meta-structure known as meta-path has been recently proposed to overcome the heterogeneity challenges. However, defining and generating good meta-paths as well as obtaining high prediction accuracy are still open issues. In this paper, a multilayered approach has been proposed in which a heterogeneous system is modeled as a multilayered complex network. Then, by exploring the network layers with different semantics, a set of meta-paths is generated. Extracting a number of topological features for each meta-path, a number of link predictors is learned which are aggregated to build the final link predictor. The experimental results on a bibliography network (DBLP) show that the proposed approach obtains higher accuracy in comparison with popular heterogeneous proximity measures.