Interpretable machine learning models for detecting fine-grained transport modes by multi-source data

Analysis of transport mode choice is crucial in transportation planning and optimization. Traditionally, the transport mode of individuals is detected by discrete choice models (DCMs), which rely on data regarding individual and household attributes. Using these attribute data raises privacy concerns and limits the applicability of the model. Meanwhile, the detection results of DCMs may be biased, despite providing insight into the impact of variables. The machine learning models are more effective for mode detection, but most models need more interpretability. In this study, an interpretable machine learning model is developed to detect the transport modes of individuals. The mobility features of individuals, which introduce the velocity and acceleration of the center of mass (COM) are innovatively considered in the detection model. These mobility features are combined with multi-source data, including land use mix, GDP, population and online map service data as detection features. Using the travel survey data from Nanjing, China in 2015, the effects of different machine learning models on fine-grained detection performance are investigated. The results indicate that the deep forest model presents the best detection performance and achieves an accuracy of 0.82 in the test dataset, demonstrating the effectiveness of the proposed detection model. Furthermore, t-distributed stochastic neighbor embedding (t-SNE) and ablation experiments are conducted to overcome the non-interpretability issue of the machine learning models. The results show that the mobility features of individuals are the most critical features for improving detection performance. This study is essential for improving the structure of transport modes and maintaining low-carbon and sustainable development in urban traffic systems.