해석 가능한 기계 학습 접근법을 이용한 부산-후쿠오카간 한일해저터널 개통에 따른 수단 선택의 선호 분석에 관한 연구

Followed by recent vibrant effort on the application of machine-based nonparametric approaches such as machine learning (ML), ML approaches have proven to be successful in many real-world applications. Despite their predictive performances, the acknowledged drawback of ML techniques is their lack of interpretable features that can be used to evaluate policy impacts. In this sense, this article aims to not only develop discrete choice modeling using ML techniques but also provide the useful insights about demystifying (i.e., interpreting) ML models. In particular, XGBoost, rule-based ensemble model combined with ML, is adopted to predict and understand mode choice behaviors related to Korea-Japan undersea tunnel high-speed train. XGBoost model notably possesses more interpretable algorithmic features than other ML models as well as high degree of predictive power for multi-level data—e.g., stated preference (SP) survey. Overall model estimates were accurated at 81% and also presented high accuracy for each model choice. To interpret users’ behaviors and preferences, variable importance (VI), variable contribution, and interactive partial dependence were estimated. The results of VI indicates that air mode, for instance, was affected by order of air fare, Fukuoka visit frequency, and Japan visit frequency, ferry mode was affected by order of Age, Willing to use tunnel, and Japan visit by air mode, and train mode was affected by train fare, age, and Japan visit frequency.