Genetic Algorithm Based Hyperparameter Optimization of XGBoost for Car Price Determination with Smooth Target Encoding and SHAP-Based Interpretability

Abstract

Car price prediction framework mostly suffers from sub-optimal parameter optimization. This study introduces an intelligent hybrid car pricing system that combines evolutionary computing with ensemble learning to construct highly precise predictions. In this study established on the noisy car price dataset, some preprocessing phases such as cleaning numerical and textual errors, handling missing data, and One-Hot Encoding for categorical features have been applied. To determine car price, XGBoost  machine learning method, efficient gradient boosting technique, has been chosen for this study. Rather than depending on the computationally expensive Grid Search, to maximize the success of the model, Genetic Algorithm with 5-Fold Cross-Validation has been used over an extended XGBoost hyperparameters. The trained models have been assessed using MAE, MAPE, RMSE, and R² metrics. In addition, error distributions and actual-versus-predicted alignments have been visualized with figures. To explain the ensemble networks, TreeSHAP (SHapley Additive exPlanations) has been introduced to mathematically quantify and visualize feature-level impacts on car price prediction. The results demonstrate that the Genetic Algorithm optimized XGBoost model provides high explanatory power and low error in car price determination.