Machine Learning (ML) approaches are increasingly being investigated as an alternative to Random Utility Models (RUMs) for modelling passenger mode choice. These approaches have the potential to provide valuable insights into choice modelling research questions. However, the research and the methodologies used are fragmented. Whilst systematic reviews on RUMs for mode choice prediction have long existed and the methods have been well scrutinised for mode choice prediction, the same is not true for ML models. To address this need, this paper conducts a systematic review of ML methodologies for modelling passenger mode choice. The review analyses the methodologies employed within each study to (a) establish the state-of-research frameworks for ML mode choice modelling and (b) identify and quantify the prevalence of methodological limitations in previous studies.

A comprehensive search methodology across the three largest online publication databases is used to identify 574 unique records. These are screened for relevance, leaving 70 peer-reviewed articles containing 73 primary studies for data extraction. The studies are reviewed in detail to extract 17 attributes covering five research questions, concerning (i) classification techniques, (ii) datasets, (iii) performance estimation, (iv) hyper-parameter selection, and (v) model analysis.

The review identifies ten common methodological limitations. Five are determined to be methodological pitfalls, which are likely to introduce bias in the estimation of model performance. The remaining five are identified as areas for improvement, which may limit the achieved performance of the models considered. A further six general limitations are identified, which highlight gaps in knowledge for future work.

越来越多地研究机器学习（ML）方法，以替代随机效用模型（RUM）对乘客模式选择进行建模。这些方法有可能为选择模型研究问题提供有价值的见解。但是，研究和使用的方法是分散的。尽管长期以来一直在对RUM进行模式选择预测的系统评价，并且已经对模式选择预测的方法进行了详尽的审查，但对于ML模型却并非如此。为了满足这一需求，本文对用于建模乘客模式选择的机器学习方法进行了系统回顾。

使用横跨三个最大的在线出版物数据库的全面搜索方法来识别574个唯一记录。筛选这些文献的相关性，留下70篇经同行评审的文章，其中包含73项主要研究用于数据提取。对研究进行了详细审查，以提取涉及五个研究问题的17个属性，这些问题涉及（i）分类技术，（ii）数据集，（iii）性能估计，（iv）超参数选择和（v）模型分析。

该评价确定了十种常见的方法学局限性。有五个被确定为方法上的陷阱，这可能会在模型性能的估计中引入偏差。其余五个确定为需要改进的领域，这可能会限制所考虑模型的实现性能。确定了另外六个一般限制，这些限制突出了将来工作中知识的空白。