An Efficient Bet-GCN Approach for Link Prediction

The task of determining whether or not a link will exist between two entities, given the current position of the network, is called link prediction. The study of predicting and analyzing links between entities in a network is emerging as one of the most interesting research areas to explore. In the field of social network analysis, finding mutual friends, predicting the friendship status between two network individuals in the near future, etc., contributes significantly to a better understanding of the underlying network dynamics. The concept has many applications in biological networks, such as finding possible connections (possible interactions) between genes and predicting protein-protein interactions. Apart from these, the concept has applications in many other areas of network science. Exploration based on Graph Neural Networks (GNNs) to accomplish such tasks is another focus that is attracting a lot of attention these days. These approaches leverage the strength of the structural information of the network along with the properties of the nodes to make efficient predictions and classifications. In this work, we propose a network centrality based approach combined with Graph Convolution Networks (GCNs) to predict the connections between network nodes. We propose an idea to select training nodes for the model based on high edge betweenness centrality, which improves the prediction accuracy of the model. The study was conducted using three benchmark networks: CORA, Citeseer, and PubMed. The prediction accuracies for these networks are: 95.08%, 95.07%, and 95.3%. The performance of the model is comprehensive and comparable to the other prior art methods and studies. Moreover, the performance of the model is evaluated with 90.13% for WikiCS and 87.7% for Amazon Product network to show the generalizability of the model. The paper discusses in detail the reason for the improved predictive ability of the model both theoretically and experimentally. Our results are generalizable and our model has the potential to provide good results for link prediction tasks in any domain.