A gravitation-based link prediction approach in social networks

Abstract Performance improvement of similarity based link prediction is an important task in social network analysis as an active research. The local, global and community information integrating can increase the prediction accuracy and the time consuming. We present a novel unsupervised gravitation-based link prediction approach to accuracy improvement of local and global predictions by integrating node features, community information and graph properties to distribute and reduce of the prediction space. The local prediction accuracy improves by proposed gravitation-based optimized subgraphs extraction from a community detection method and reduces the global prediction by search space distribution to network capacity increasing. The approach is demonstrated to be more accurate, adaptive and scalable than some existing similarity index-based methods with a significant reduction of the running time through experimental results. The accuracy improves with more existence powerful communities, triangle links and low diameter. As a trade-off between accuracy and execution time, the approach may also be applicable to large and complex networks.