Predicting Engagement in Human-Robot Teams via Node-Edge Co-Attention Dynamic Graph Neural Networks

Abstract

Human-AI collaboration increasingly depends on intelligent agents that can understand and influence human social behavior. In small-group settings, conversational dynamics shape team cohesion and task success. However, existing models in dynamic graph learning struggle with small-scale graphs, high-dimensional edge features, and multi-level predictions. This paper proposes a novel Node-Edge Co-Attention Dynamic Graph Neural Network (DyNEA) for engagement prediction in human-robot teams. Using data from 30 rounds of collaborative games involving conversations from participants, our model jointly learns node, edge, and graph-level representations and make predictions. DyNEA outperforms baselines across multiple metrics. Our framework offers potential applications in human-AI collaboration, emotional support systems, and modeling cooperation dynamics.