Predicting visual aesthetic preferences in Tehran city universities campuses using machine learning techniques

Visual aesthetic preferences fundamentally shape the restorative potential of university landscapes and have a significant impact on student well-being and engagement. This study developed Ensemble Learning Models to predict students' aesthetic preferences for interactive rest spots and compared their accuracy with conventional individual models. The input dataset (18 features) was extracted from images of 100 student rest spots across four universities campuses in Tehran city: University of Tehran, Amirkabir University of Technology, Shahid Rajaee Teacher Training University, and Tarbiat Modares University. Based on the aesthetic preferences reported by 394 university students, the study employed Support Vector Regression (SVR), Random Forest (RF), Multilayer Perceptron (MLP), and their combinations of SVR-MLP and SVR-RF-MLP to predict the aesthetic quality on university campuses. The results show that Ensemble Learning Models outperform individual models in predicting students' aesthetic preferences, filling a key research gap. The individual models demonstrated varying levels of accuracy across the total dataset, with SVR (R² = 0.824) performing the strongest, followed by MLP (R² = 0.814) and RF (R² = 0.761). Among all, the SVR-MLP ensemble learning model achieved the highest accuracy, with R² scores of 0.767 (test data), 0.850 (training data), and 0.828 (total dataset). Key design elements enhancing both aesthetic appeal and mental restoration included more trees, soft landscapes, waterscapes, and color diversity, coupled with minimal building and pathway presence. The Ensemble Learning Models provide a robust conceptual framework for architects, environmental designers, landscape architects, and campus planners to design attractive and restorative spaces aligned with students' visual preferences.