A MATHEMATICAL ORDINAL MULTICLASS FRAMEWORK FOR ELECTRIC-VEHICLE ADOPTION INTENTION BASED ON BAYESIAN-OPTIMIZED XGBOOST

Abstract

As electric-vehicle markets broaden, consumers do not enter with a single, uniform level of adoption intention. This study asks whether Low, Medium, and High intention can be separated reliably in an imbalanced sample and which observed variables do most of the separating. The data contain 50,000 consumers and 21 measures spanning demographics, mobility, charging, cognition, prior experience, and cost. Cleaning, encoding, and standardization produce 25 model inputs. Spearman correlation is used to inspect monotone relationships before a regularized multiclass XGBoost model is fitted with a stratified 80:20 split, stratified five-fold cross-validation, and 100 Bayesian optimization trials. Gain and SHAP are then used to examine overall and class-specific contributions. Mean cross-validation accuracy is 86.88% (SD 0.61%), and accuracy on the independent test set is 87.40%. F1-scores are 0.94 for High, 0.83 for Low, and 0.74 for Medium. The weaker Medium result reflects overlapping distributions and greater within-class variation. Electric-vehicle knowledge and range anxiety are the leading monotone correlates and together account for 0.419 of Gain importance. The findings connect predictive performance with boundary uncertainty and feature contribution in an interpretable ordinal-classification setting.