林波

港珠澳大桥的两端西面连接着珠海市区和澳门人工岛，东面止于香港国际机场东部的香港口岸人工岛，建成连接香港、珠海和澳门三地间的交通要道。面对跨境交通出行需求的不断增加，居民对跨境智慧出行服务提出了更高的要求，因此明晰跨境出行人群出行方式选择意向产生的内在机理是进行跨境交通管理与控制的重要工作。故本文旨在分析港珠澳大桥出行者的跨境出行方式选择影响因素，以及预测多模式跨境出行行为，为智慧跨境交通出行系统的发展与结构调整提供可靠的理论依据。

首先，本文利用网络问卷调查数据收集了港珠澳大桥跨境出行者Revealed Preference（RP）实际出行信息、Stated Preference（SP）出行意愿信息以及个人社会经济属性三类信息。对问卷调查资料进行筛选、编码录入以及问卷信度分析和效度检验后，再对RP/SP样本的信息做出统计性描述，分析港珠澳大桥跨境出行者的人口特征、实际出行规律以及出行偏好特征。

然后，跟据有效问卷样本数据中的观测变量建立结构方程模型，以获取跨境出行者的潜变量估计值，将其和显变量中的数据进行数据融合，建立经过贝叶斯模型平均算法（Bayesian Model Averaging, BMA）方法二次建模处理的Mixed Logit模型，并对整合模型进行参数估计。模型结果表明：影响港珠澳大桥跨境出行方式选择行为的主要显著性因素有性别、出行目的（旅游）、出行起点至港珠澳大桥珠海口岸的交通方式（公交和私家车）、港珠澳大桥香港口岸至出行目的地的交通方式（公交、私家车、出租车和地铁）、出行时间（港珠澳大桥香港口岸至出行目的地）和换乘次数，SP出行情景变量中的出行时间、出行花费、等候时间和情景状态依赖变量，以及安全性，舒适性，便捷性三个潜变量属性。与未经过BMA方法处理的Mixed Logit模型进行对比分析，其结果验证了整合模型在模型解释能力以及拟合效果上的优越性。因此，可以应用提出的整合模型对港珠澳大桥跨境出行行为进行建模分析，有利于解释跨境出行人群出行选择意向产生的内在机理。

最后，论文围绕如何更好地为跨境出行者提供一体化出行服务，提高多模式跨境出行类别不平衡数据的分类预测性能展开，引入了边界合成少数类过采样算法（Borderline Synthetic Minority Over-sampling Technique，BLSMOTE）算法对出行数据集进行类别标签数据平衡化处理，构建了随机森林模型对跨境出行模式进行分类预测，然后采用SHAP值对预测模型的影响特征进行解释分析，最后与多种机器学习算法进行模型性能的比较。研究结果表明BLSMOTE算法可以有效提高不平衡数据集分类预测的精度，与随机森林模型结合的整合模型精度更高，能够用来预测跨境出行者多模式跨境出行情况，有利于向出行者推荐定制化的跨境智慧出行模式方案。

关键词：跨境出行，Mixed Logit模型，贝叶斯模型平均法，BLSMOTE算法，随机森林模型

The Hong Kong-Zhuhai-Macao Bridge (HZMB) starts from the Zhuhai and Macau Artificial Islands in the west, and ends at the Hong Kong Port Artificial Island in the east. It is a major traffic road connecting Hong Kong, Zhuhai and Macau. With the increasing demand for cross-border traffic, residents have put forward higher requirements for smart travel services. Therefore, it is important to carry out cross-border traffic management and control to clarify the mechanism of cross-border travel choice. This paper aims to analyze the influencing factors of cross-border travel mode choice and to predict multi-mode cross-border travel behaviors, and the results can provide a reliable theoretical basis for the development of smart cross-border travel systems.

Firstly, three types of information, including Revealed Preference (RP) actual travel information, Stated Preference (SP) travel willingness information, and personal socio-economic attributes of cross-border travelers on the HZMB were collected by online questionnaire in this paper. After screening, coding, and reliability analysis and validity test of the questionnaire data, the RP/SP sample data were statistically described. Additionally, the demographic characteristics, actual travel patterns and travel preference characteristics of cross-border travelers on the HZMB were analyzed. Therefore, the proposed integrated model can be used to model and analyze the cross-border travel behavior of the HZMB, which is beneficial to explain the internal mechanism of the cross-border travel choice intention.

Secondly, according to the observed variable in the questionnaire, a structural equation model is constructed to obtain the estimated value of latent variables. And the sample data is obtained with the explicit variable and latent variables. The Mixed Logit model processed by the BMA method is used to analyze the RP and SP fusion data. The integrated model estimation results show that the main significant factors affecting the choice of cross-border travel mode of the HZMB are gender, tourism, transportation mode (bus and private car) from the departure point to the Zhuhai Port of the HZMB, transportation modes (bus, private car, taxi and subway) on the HZMB, travel time (the Hong Kong port to destination) , number of transfers, travel time, travel cost, waiting time in SP travel scenario, state dependent variables, and three latent variable including safety, comfort, and convenience. Compared with the Mixed Logit model, the results verified the superiority of the integrated model in terms of model interpretation ability and goodness of fit.

Finally, the paper validated that overlooking the characteristics of multi-mode cross-border travel imbalance data will lead to a decline in classification prediction performance. Thus, the BLSMOTE algorithm is proposed to balance the data. And random forest model was built to classify and predict the cross-border travel mode based on the travel balanced data set. Then the SHAP value is used to explain and analyze the influence characteristics of the prediction model. Compared with multiple machine learning algorithms, the research results show that the BLSMOTE algorithm can effectively improve the accuracy of classification and prediction of imbalanced datasets. And the random forest model with the BLSMOTE algorithm can be used to predict the multi-modal cross-border travel situation of cross-border travelers, which is conducive to recommending customized travel mode to travelers.

Key Words: Cross-border travel, Mixed Logit model, Bayesian model averaging method, BLSMOTE algorithm, Random forest