Persona-based stratification and XGBoost modeling for multi-dimensional vehicle sound quality evaluation

As the automotive industry transitions toward electrification and intelligent mobility, conventional NVH development approaches struggle to meet users' personalized sound quality (SQ) expectations. This study systematically investigates the impact of individual differences on multi-dimensional SQ evaluation and develops a robust explanatory reconstruction framework. Using a bipolar 7-point scale, 50 participants provided ratings along the Comfort and Sportiness dimensions, and a threshold-based preference split was applied to construct detailed user personas. Extreme Gradient Boosting (XGBoost) was employed to model the nonlinear mapping between psychoacoustic features and subjective ratings, with SHAP (SHapley Additive exPlanations) analysis applied to quantify the relative contributions of individual features to each perceptual dimension. The analysis focuses on high-fidelity full-sample models for feature decoupling. SHAP analysis further demonstrates that the derived feature contributions are robust and interpretable across perceptual dimensions. These findings confirm that persona-based stratification allows investigation of population heterogeneity in a mechanistic perspective, providing a physically interpretable basis for precise vehicle SQ tuning.