Inferring the internal structure of social collectives

We investigate how humans infer the rich internal structure of social collectives from patterns of interactions between agents. We propose a computational model of this process which integrates a domain-general statistical learning mechanism with, domain-specific knowledge about social contexts (i.e.: "intuitive sociologies"). We test our model in two experiments where participants observe a sequence of animated interactions between agents, and then assign the agents to groups according to their role or type within the social collective. Crucially, the two experiments depict different types of social interactions which reflect different types of underlying social structures. The patterns of correspondence between model predictions and human data support our account, and demonstrate the importance of both general statistical learning and specific social knowledge when reasoning about social collectives.