Explainable AI-Based Identification of Floor Plan Characteristics Influencing the Seismic Performance of Wooden Houses

Understanding how floor plan characteristics influence seismic performance is essential for achieving both safety and design flexibility in wooden houses. This study quantitatively identifies the relative importance of floor plan parameters affecting seismic response deformation in two-story wooden houses in Japan using machine learning and explainable artificial intelligence (XAI). Analytical models were constructed from 60 actual houses, generating 840 seismic response cases by varying ground-motion levels. Building-level five-fold grouped cross-validation was performed to evaluate generalizability to unseen buildings, yielding an R2 of 0.645±0.158 and an RMSE of 16.869±3.538 (mm). Because the primary objective was to characterize relationships within the present parametric dataset, LightGBM, which achieved the highest accuracy under the case-level random split (R2=0.8993, RMSE = 9.1888 (mm)), was used for detailed SHAP-based interpretation. Grouped-fold SHAP analysis further indicated that the principal feature-importance rankings were broadly stable. Wall alignment ratio (W) and floor area ratio (Ra) were identified as influential structural parameters. Exploratory inspection of the SHAP dependence plots suggested possible changes in the dependence trends around W ≈ 0.45 and WW2 ≈ 0.75; however, these values were not validated as stable transition values across the grouped folds and should be regarded as dataset-specific observations. The results demonstrate the potential of XAI for quantitatively interpreting seismic-response mechanisms in wooden houses.