探討機器學習法的預測性與解釋性：以MaaS方案的選擇行為為例

交通行動服務 (Mobility as a Service, MaaS) 是一種交通資源整合的新穎概念，過去研究多著重應用個體選擇模式進行分析，而現今機器學習因自動化和高效精確的預測能力，已成為靈活且彈性的預測分析工具。另，當今研究鮮少針對敘述性偏好數據進行實證研究，變數型態與效用函數指定的探討亦是本研究的探討重點。因此本研究藉由MaaS方案選擇的敘述性偏好數據，進行機器學習與混合羅吉特模式的比較，以混淆矩陣與預測正確百分比作為預測性指標，並透過變數重要性與T值顯著性探討解釋性的變數重要程度，以及部分相依圖與彈性分析探討解釋性的變數影響程度。研究結果顯示在含有敘述性偏好的交叉變數為「方案價格/月入」時的訓練成效最佳，並證實機器學習的預測性優於混合羅吉特模式，其中，以Bagging的隨機森林與Boosting的梯度提升決策樹為最佳之集成式演算法。在變數重要與影響程度的解釋性中，研究顯示發現機器學習與個體選擇模式無法客觀地直接比較各自優勢，但透過外部程序的整合方式，可增強模式的預測性與解釋性。研究結果顯示在一般大眾族群中模式間的變數重要排序相同；在年輕學生族群中變數重要排序雖有不同，但可證實機器學習能有效輔助混合羅吉特模式進行旅運行為與社經變數的效用函數指定；然，因不同地區的變數選取不盡相同，故在地域性和族群性的差異考量下，透過機器學習的重要變數之先行分析，可有效作為個體選擇模式的效用函數之指定依據，從而結合機器學習的預測優勢與個體選擇模式的解釋能力。

Mobility as a Service (MaaS) is an innovative concept that integrates various trans-portation resources. While previous research has predominantly focused on using discrete choice models for analysis, machine learning has emerged as a flexible and adaptive pre-dictive analysis tool due to its automation and highly accurate forecasting capabilities. Moreover, current research seldom focuses on empirical studies using stated preference data. The study emphasizes the exploration of variable types and utility function specifica-tions. Therefore, this research compares machine learning with the mixed logit model us-ing stated preference data for MaaS. Predictive accuracy is assessed through confusion matrices and correct prediction percentages. Furthermore, variable importance and T-value significance are used to assess the importance of explanatory variables, while partial dependence plots and elasticity analysis examine their impact. Results indicate that train-ing performance is optimal when the cross variable "price of alternative/monthly income" is included, and machine learning demonstrates superior predictive capability over the mixed logit model. The best integrated algorithms are Bagging's random forest and Boost-ing's gradient boosting decision tree. The research indicates that it is difficult to objective-ly compare the strengths of machine learning and discrete choice models in explaining variable importance and impact. However, integrating external procedures can enhance both the predictive and explanatory capabilities of the models. The research indicates that variable importance rankings are consistent across different models for the general popula-tion. However, among the student population, the rankings differ. This indicates that ma-chine learning can effectively complement the mixed logit model in specifying utility functions for travel behavior and socio-economic variables. Due to regional and demo-graphic differences in variable selection, preliminary analysis of key variables using ma-chine learning can effectively inform the specification of utility functions in discrete choice models. This approach combines the predictive advantages of machine learning with the explanatory of discrete choice models.