基于机器学习的时空出行选择行为研究综述与展望  

A review and outlook of machine learning-based travel choice behavior research

作　　者：王江波[1] 连芝锐 冯涛[2] 唐立 刘锴[1] WANG Jiangbo;LIAN Zhirui;FENG Tao;TANG Li;LIU Kai(School of Economics and Management,Dalian University of Technology,Dalian 116024,Liaoning,China;Graduate School of Advanced Science and Engineering,Hiroshima University,Higashi-Hiroshima 739-8511,Japan;School of Automobile and Transportation,Xihua University,Chengdu 610031,China)

机构地区：[1]大连理工大学经济管理学院,辽宁大连116024 [2]广岛大学先进理工系科学研究科,日本东广岛739-8511 [3]西华大学汽车与交通学院,成都610031

出　　处：《地理科学进展》2024年第8期1649-1665,共17页Progress in Geography

基　　金：国家自然科学基金项目(52302404,52102384,71871043);宁夏回族自治区揭榜挂帅重点项目(2023BBF01004)。

摘　　要：近年来,机器学习模型因其优越的预测性能和灵活性,被广泛引入时空出行行为建模与预测研究中,但其基础研究框架和技术路线尚未明晰。论文通过回顾2010—2022年相关领域发表的重要文献,梳理了机器学习算法的应用对时空出行选择行为研究范式的影响,总结了当前研究中亟待解决的关键问题及影响时空出行选择行为建模效果的潜在因素和作用机理,展望了未来研究中需要重点突破的方向。将机器学习算法有效应用于时空出行选择行为研究,不仅需要与决策场景相契合的模型架构和决策机理支撑,还应克服诸多机器学习过程及方法的固有缺陷,并充分考虑外部研究条件对时空选择行为模拟和预测效果的影响。现有的机器学习模型已能够契合大多数出行选择决策场景,多元化、高效率的机器学习算法必将有力推动出行选择行为研究的发展。有限的可解释性仍然是学者们难以广泛信任基于机器学习的时空出行选择行为模型的根本原因。面对大数据时代时空出行选择行为研究的机遇与挑战,充分融合机器学习算法和经典决策理论及模型各自的优势,同时提升时空出行选择行为的模拟精度和模型可解释性是重要发展趋势。In recent years,machine learning models have been widely introduced into spatiotemporal travel behavior modeling and prediction research due to their superior predictive performance and flexibility,but their underlying research framework and technical routes are still unclear.This article reviewed the typical literature published in related fields from 2010 to 2022 to examine the impact of the application of machine learning algorithms on the spatiotemporal travel choice behavior research paradigm,summarize the key issues to be solved in the current application and the potential influencing factors and mechanisms that affect the effectiveness of spatiotemporal travel choice behavior modeling,and foresee the directions to be focused on in future research.The effective application of machine learning algorithms to the study of spatiotemporal travel choice behavior requires not only the support of model architectures and decision mechanisms that fit the decision scenarios,but also to overcome the inherent shortcomings of all learning processes and methods,and fully consider the impact of external research conditions on the simulation and prediction performance of spatiotemporal travel choice behavior.Existing machine learning models can already fit most spatiotemporal travel choice decision scenarios,and diversified and efficient machine learning algorithms will certainly give a strong impetus to the development of spatiotemporal travel choice behavior research.Limited model interpretability remains the fundamental reason why machine learning-based spatiotemporal travel behavior models are difficult to be widely trusted.Facing the opportunities and challenges of spatiotemporal travel choice behavior research in the era of big data,it will be an important development trend to fully integrate the respective advantages of machine learning algorithms and classical decision theories and models,while improving the simulation accuracy and model interpretability of spatiotemporal travel choice behavior.

关 键 词：时空出行选择行为 非集计模型 机器学习 可解释性 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] U491[自动化与计算机技术—控制科学与工程]