Predictive Role Discovery of Research Teams Using Ordinal Factorization Machines

Abstract

In this paper, we address the problem of research role discovery, especially for large research institutes where similar yet separated teams co-exist. The roles that researchers play in a research team, i.e., principal investigator, sub-investigator and research staff, typically exhibit an ordinal relationship. In order to better incorporate the ordinal relationship into a role discovery model, we approach research role discovery as an ordinal regression problem. In the proposed approach, we represent a research team as a heterogeneous teamwork network and propose OrdinalFM, short for Ordinal Factorization Machines, to learn the role prediction function. OrdinalFM extends the traditional Factorization Machines (FM) in an effort to handle the ordinal relationship among learning targets. Experiments with a real-world research team dataset verify the advantages of OrdinalFM over state-of-the-art ordinal regression methods.