Ensemble Link Predictor for Heterogeneous Complex Networks

Link prediction in complex networks aims to explore similarities between node pairs. Recently, link prediction has been considered in the presence of network heterogeneity which makes the majority of the homogeneous link prediction approaches infeasible. A meta-structure, known as meta-path, has been proposed to explore such networks. Generating good meta-paths and selecting the best of them introduce some new challenges to link prediction problem. In this paper, a new ensemble-based link prediction approach is proposed in heterogeneous complex networks. This approach consists of three steps: (i) a set of meta-paths are selected such that each of them represents a different semantic between the target node pairs; (ii) a feature vector is extracted for each node pair using each meta-path; (iii) an ensemble of learners would be established on different feature sets. The final link predictor is obtained after the ensemble aggregation. The results on DBLP network show that the proposed approach has more accurate predictions than a single meta-path based link predictor.