An interpretable prediction model for pavement performance prediction based on XGBoost and SHAP

Abstract

Considering traditional methods' low accuracy and machine learning methods' lack of interpretability, this paper proposed a pavement performance model for IRI prediction based on XGBoost, and introduced SHAP to enhance the interpretability of individual features of the model. The data used are from the America LTPP data. Firstly, data cleaning and preprocessing were conducted. Secondly, 4 prediction models were built based on classical algorithms, namely, LightGBM, XGBoost, SVM, and multiple linear regression. Then, by comparison, it was found that XGBoost performed better. Finally, parameter tuning for this model was performed, with the RMSE as 0.317, MAE as 0.219, and R 2 as 0.742. In addition, considering the prediction model's lack of transparency, SHAP is utilized to perform the feature importance analysis and identify the main factors affecting the pavement performance, which can help the highway sector to improve the reliability of their subsequent prediction model analysis.