Automatic role identification for research teams with ranking multi-view machines

Abstract

Research teams have been well recognized as the norm in modern scientific discovery. Rather than a loose collection of researchers, a well-performing research team is composed of a number of researchers, each of them playing a particular role (i.e., principal investigator, sub-investigator or research staff) for a short- or long-term effort. Role analysis for research teams would help gain insights into the dynamics of teams and the behavior of their members. In this paper, we address the problem of research role identification for large research institutes in which similar yet separated teams coexist. In particular, we represent a research team as teamwork networks and generate the feature representation of each member using a number of network metrics. Afterward, we propose RankMVM, short for Ranking Multi-View Machines, to learn the role identification model. Compared with traditional predictive models, RankMVM is advantageous in exploring high-order feature interactions in an efficient way, as well as handling the specific challenges of the research role identification task, including partially ordered learning targets and sparse feature interactions. In the experiments, we assess the performance on a real-world research team dataset. Extensive experimental evaluations verify the advantages of our proposed research role identification approach.