Data Sheet 1_Classifying interpersonal interaction in virtual reality: sensor-based analysis of human interaction with pre-recorded avatars.pdf

This study investigates human engagement with a non-responsive, pre-recorded avatar in VR environments. Rather than bidirectional collaboration, we focus on unidirectional synchrony from human participants to the avatar and evaluate its detectability using sensor-based machine learning. Using a random forest model, we classified interactions into cooperation, conformity, and competition, achieving an F1 score of 0.89. Feature importance analysis identified hand rotation and head position as key predictors of interaction states. We compared human-human and human interaction with a non-responsive avatar (pre-recorded motion replay) during a joint Simon task by covertly switching collaborators between humans and non-responsive avatars. Using the classification model, a synchrony index was derived from VR motion data to quantify behavioral coordination patterns during joint actions. The classification indexes were associated with higher cooperation in human-human interactions (p=0.0262) and greater conformity in human interaction with a non-responsive avatar (p=0.0034). The synchrony index was significantly lower in the non-responsive avatar condition (p<0.001), indicating reduced interpersonal synchrony with non-responsive avatars. These findings demonstrate the feasibility of using VR sensor data and machine learning to quantify social interaction dynamics. This study aimed to explore the feasibility of a sensor-based machine learning model for classifying interpersonal interactions in VR, based on preliminary data from small-sample experiments.