Link prediction in complex networks using node centrality and light gradient boosting machine

Link prediction is amongst the most crucial tasks in network science and graph data analytics. Given the snapshot of a network at a particular instance of time, the study of link prediction pertains to predicting possible future links amongst the nodes of the networks. It finds applications in recommender systems, traffic prediction in networks, biological interactions, and many others. In this paper, we propose a novel generic approach to link prediction based on using various node centralities and different machine learning classifiers. We utilize some popular and recently introduced node centralities to capture better the network’s local, quasi-local and global structure. The value of various node centralities acts as the feature labels for the nodes in the network. The existent and non-existent edge in the network is labeled as positive and negative samples, respectively. The features of the nodes at the end of the edges, along with the positive or negative label, form a well-defined dataset for the task of link prediction. The dataset is then fed into various machine learning classifiers, and the best results are obtained with Light Gradient Boosted Machine (LGBM) classifier. We investigate the performance of the proposed model on multiple real-life networks using various performance metrics and reveal that our approach outperforms many popular and recently proposed link prediction methods.