Machine learning models in urban transport applications: understanding and actuation

• Main Author: Li, Mingqian
• Other Authors: Mo Li
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/171580
• Institution: Nanyang Technological University

The prosperity of big data has boosted the development of AI techniques in smart city. In this thesis, we study the adaptation of newly-emerging machine learning techniques to urban transport applications. We review existing tasks and techniques in both urban transport understanding and urban transport actuation, based on which we identify remaining challenges, which serve as the motivation of my PhD studies. My research during PhD studies primarily tackles two major challenges (in particular contexts), namely, the presence of non-recurrent mobility patterns and the experimental limitations in actuation tasks. In the first work, we tackle the challenge of non-recurrent mobility patterns in the context of highway vehicular traffic flow prediction. We adopt spatiotemporal modeling techniques and propose to mine the transition patterns from historical trajectories to complement non-recurrent mobility patterns. Specifically, we devise a model, Trajectory-based Graph Neural Networks (TrGNN), that incorporates trajectory transition patterns into the spatiotemporal deep learning framework based on graph to improve the accuracy of traffic flow prediction. Experiments with our approach on a real-world dataset achieves significant improvement on prediction accuracy, especially in non-recurrent scenarios. Our second work is motivated from the challenge of an experimental limitation encountered in the task of driver recipient selection for traffic safety education to reduce traffic accidents. We formulate the task into an uplift modeling problem, a typical problem in causal inference. We identify the challenge that due to the infeasibility of proper experiments (i.e. Randomized Control Trials) in such a real-world scenario, datasets often come with bias, which deteriorates the estimation of uplifts by existing models and evaluation metrics. This is a common challenge in uplift modeling across various domains (even beyond urban transport). We systematically study uplift modeling approaches and propose organic integration of sample re-weighting in existing uplift models and evaluation metrics. Extensive experiments with three real-world datasets as well as the case study on traffic safety education show significant performance gain from our proposed approach. A large portion of our research is found to be easily applicable to domains beyond urban transport. For example, we realize that the techniques for spatiotemporal modeling studied in our first work could be easily transferred to domains such as weather forecasting. Besides, the advance of sample re-weighting proposed in our second work can be applied to medical study with satisfactory results, as already demonstrated by our experiments with the infant dataset and the twins dataset, and can be further applied to domains like e-commerce marketing. Moreover, the major challenges we have identified in urban transport understanding and actuation, such as multi-source data fusion or knowledge transferability, can in fact be seen in various smart city applications beyond urban transport (e.g. surveillance, smart lighting, air quality management), and even domains beyond smart city. Hence, effective solutions to these challenges would bring benefits more than we expected.