INTERPRETABILITY OF MACHINE-LEARNING BASED TRAVEL BEHAVIOUR MODELS USING EXPLAINABLE AI (XAI)

Abstract

Discrete choice models are a useful method for understanding human choice behavior based on random utility theory, and have been used in various fields including transportation planning. However, discrete choice models emphasize theory and interpretability, which limits their ability to construct highly accurate models. In this study, we attempt to construct a forecasting method that excels in both “accuracy” and “interpretability” by improving forecasting accuracy with the aid of machine learning and by using an interpretative index that indicates the model’s forecasting basis. Specifically, we constructed a transportation mode selection model using a neural network, and attempted to construct a model that is interpretable and has superior forecasting accuracy by applying interpretative indices such as PD and SHAP. As a result, the model has higher prediction accuracy and more detailed interpretability than conventional models, and is expected to contribute to the advancement of policy evaluation.