Autoencoder Network for the Identification of Emotional Psychophysiological Responses in VR Urban Exploration

Population-level analyses of autonomic responses to emotional stimuli implicitly assume homogeneous physiological rules, obscuring meaningful inter-individual heterogeneity. We propose an unsupervised autoencoder trained on HRV and EDA signals from 222 subjects to identify latent autonomic response patterns without predefined labels. Spectral clustering on reconstruction correlations identified two distinct subgroups, validated on an independent virtual reality urban walk dataset (69 participants). Cluster-stratified analyses revealed systematic between-cluster differences and subgroup-specific emotional sensitivity that were invisible at the aggregate level, demonstrating that latent physiological heterogeneity can both mask true effects and generate spurious ones. The framework supports subgroup-specific characterization of autonomic responses with applications in environmental psychophysiology and affective computing.