Study of attention mechanisms for computational models of cohesion

Abstract

Social Signal Processing is a new interdisciplinary domain of research whose end goal is to equip machines with social intelligence, a facet of intelligence only known to mankind. As life happens in groups, it is essential that machines understand how human groups function to be able to collaborate with them in the future. One important aspect of group life is group cohesion. In the social sciences literature, several definitions of cohesion exist. In our work, we use Severt and Estrada’s integrative framework of cohesion. Recently, Maman et al. developed a computational model of cohesion called fItG that automatically predict the dynamics of cohesion. However, this model uses all interaction timestamps equally, without dynamically giving each of them a certain importance. In this work, we develop attention-based versions of fItG. We compare their performances with the original fItG model through permutation tests of their F1-scores