Uncovering mechanisms of co-authorship evolution by multirelations-based link prediction

Abstract

A single mechanism is insufficient for providing a comprehensive understanding of co-authorship formation and evolution because people choose to co-author with diverse motivations. The ways in which a hybrid mechanism jointly influences co-authorship evolution is not yet very clear, which leads to the following research questions: (1) how does each mechanism leverage with each other and how can multiple mechanisms be combined into the best hybrid mechanism? (2) How can the mechanisms be categorized into different groups and how does each group contribute to co-authorship evolution? This paper addresses these questions by using an improved meta-path based model called multirelations-based link prediction, which denotes every mechanism and their combinations as predictors in heterogeneous networks and quantitatively evaluates predictors via link prediction. Experiments are conducted in Library and Information Science (LIS). The result shows that the most appropriate mechanism is a hybrid mechanism denoted by a combination of predictors with weights. In addition, the contributions of different categorized mechanisms are compared, where the author-based mechanisms are more important than the keyword-based and journal-based mechanisms. The result also indicates that there is information loss when projecting from a heterogeneous bibliographic network to a homogeneous co-authorship network. Our study could add more predictive information into the model and apply the method in other types of heterogeneous networks in the future.