Graph-based analysis of work-from-home layouts and affective responses in virtual reality

This study examines a graph-based learning framework for modelling associations between work-from-home (WFH) furniture layouts and users' affective responses in immersive virtual reality (VR). Forty-three participants evaluated two layout conditions-self-arranged and researcher-arranged-in a controlled 4.2 × 4.2 m room environment, yielding 85 valid layout cases. Layouts were reconstructed in Unity and rated on six outcomes: perceived visual complexity, satisfaction, concentration, sense of control, inclination to use the workspace, and perceived performance. Each layout was represented as a spatial graph, with furniture and architectural elements encoded as nodes and proximity-based relationships encoded as edges. A GraphSAGE model with attention pooling jointly predicted the six outcomes under weighted MSE, achieving a mean absolute error of 0.684 and variance-weighted R² of approximately 0.659 in five-fold cross-validation. Compared with non-graph baseline models, the graph-based model showed lower prediction error in this dataset while also providing interpretable node- and edge-level importance patterns. GNNExplainer highlighted desk-related relationships, including desk-window, desk-door, and desk-storage configurations, as influential subgraphs. Rather than establishing a general theory of space-emotion interaction, the study demonstrates the methodological value of graph-based representation for evaluating how object-level spatial relationships in WFH layouts are associated with user experience.