A comparative analysis of local similarity metrics and machine learning approaches: application to link prediction in author citation networks

Understanding the evolution of paper and author citations is of paramount importance for the design of research policies and evaluation criteria that can promote and accelerate scientific discoveries. Recently many studies on the evolution of science have been conducted in the context of the emergent Science of Science field. While many studies have probed the link problem in citation networks, only a few works have analyzed the temporal nature of link prediction in author citation networks. In this study we compared the performance of 10 well-known local network similarity measurements with four machine learning models to predict future links in author citations networks. Differently from traditional link prediction methods, the temporal nature of the predict links is relevant for our approach. Our analysis revealed that the Jaccard coefficient was found to be among the most relevant measurements. The preferential attachment measurement, conversely, displayed the worst performance. We also found that the extension of local measurements to their weighted version do not significantly improved the performance of predicting citations. Finally, we also found that a XGBoost and neural network approach summarizing the information from all 10 considered similarity measurements was able to provide the highest AUC performance and competitive precision values.