Using Explainable AI to Understand Team Formation and Team Impact

Abstract

ABSTRACT The citation of scientific papers is considered a simple and direct indicator of papers' impact. This paper predicts papers' citations through team‐related variables, team composition, and team structure. Team composition includes team size, male/female dominance, academia/industry collaboration, unique race number, and unique country number. Team structures are made up of team power level and team power hierarchy. Team members' previous citation number, H‐index, previous collaborators, career age, and previous paper numbers are a proxy of team power. We calculated the mean value and Gini coefficient to represent team power level (the collective team capability) and team power hierarchy (the vertical difference of power distribution within a team). Taking 1,675,035 CS teams in the DBLP dataset, we trained the XGBoost model to predict high/low citation. Our model has reached 0.71 in AUC and 70.45% in accuracy rate. Utilizing Explainable AI method SHAP to evaluate features' relative importance in predicting team citation categories, we found that team structure plays a more critical role than team composition in predicting team citation. High team power level, flat team power structure, diverse race background, large team, collaboration with industry, and male‐dominated teams can bring higher team citations. Our project can provide insights into how to form the best scientific teams and maximize team impact from team composition and team structure.