Explainable Machine Learning and Generative AI for SHAP-Guided Modification of Residential Interior Images: Predictive Associations and Algorithm-Derived Spatial-Quality Scores

With the increasing adoption of data-driven methods in architecture and design, the quantitative assessment of residential interior spatial quality and model-informed modification of residential interior images have become increasingly important. This study develops an explainable machine-learning and generative AI-assisted framework linking spatial-quality assessment, design-variable diagnosis, model interpretation, and design modification. Based on 13,962 residential interior images, 12 image-derived design variables were extracted using OpenCV and YOLO. Functionality score (FS), healthiness score (HS), aesthetics score (AS), and comprehensive score (CS) were used as supervised learning targets, and Random Forest (RF), GBDT, XGBoost, and LightGBM models were compared and optimized. RF achieved the best overall performance, with held-out test-set R2 values ranging from 0.3995 to 0.6685. SHAP analysis identified distinct contribution patterns and nonlinear associations among brightness, furniture configuration, plant-related variables, center blank ratio, openness, and color characteristics. A sensitivity analysis further showed that the predictive performance of HS and CS decreased after removing predictors with potential computational overlap with the previous scoring framework. These results warrant cautious interpretation of the corresponding associations. In three representative cases, SHAP-derived patterns were translated into AIGC-assisted design strategies, and the mean recalculated algorithm-derived CS increased from 0.2751 to 0.5673, while the decrease in AS indicated potential trade-offs among multiple spatial-quality objectives. The proposed framework provides an interpretable computational pathway for image-based residential interior assessment and model-informed design modification evaluated through changes in algorithm-derived spatial-quality scores.