The capacity to predict human spatial preferences within built environments is instrumental for developing Cyber-Physical-Social Infrastructure Systems (CPSIS). A significant challenge in this domain is the generalizability of preference models, particularly their efficacy in predicting preferences within environmental configurations not encountered during training. While deep learning models have shown promise in learning complex spatial and contextual dependencies, it remains unclear which neu...