Computational Approaches to Team Modeling: Insights from Team Potency

This thesis explores the intersection of social and computational team analysis, aiming to raise awareness around the theoretical and conceptual frameworks that underpin team analysis since it inception. Developing theory-informed computational models of team phenomena is a pivotal requirement for designing socially intelligent machines that can be integrated into hybrid human-machine teams. To achieve this ambitious goal, this thesis focuses on team potency -- a team emergent phenomenon representing the collective belief in a team that it can be effective. To the best of our knowledge, this thesis is the first of its kind to propose an automated approach to modeling team potency from the perspective of an external observer, relying on behavioral cues. Investigating the team potency nature, this work addresses core modeling challenges within the field of computational team analysis, contributing two graph-based modeling approaches along with two Graph Neural Networks (GNN) frameworks. These frameworks aim to serve as backbone modeling strategies for tackling key challenges in team analysis, particularly in modeling (i) the multilevel nature of teams and (ii) the dynamic team composition, where teams may vary in size. Furthermore, to alleviate the burden of manual data annotation, this thesis proposes a novel GNN-based approach that propagates information from a small subset of labeled nodes to a larger pool of unlabeled nodes, thereby facilitating more efficient data annotation for team analysis. Beyond these general contributions to computational team analysis, the thesis specifically addresses challenges in modeling team potency. Particularly, it challenges the traditional aggregation strategy employed to derive team-level properties from lower-level attributes. Overall, the findings emphasize the importance of incorporating social science theories into the design of computational models of team phenomena. More broadly, this work underscores the challenges of grounding high-level conceptual theories in low-level computational models based on automated behavioral cues. Thus, calling for a more integrative approach to team analysis and prompting for collaborative research between social and computer scientist to address joint research questions.