Temporal and Structural Machine Learning from Transportation Data

Open Access
Zhan, Hongyuan
Graduate Program:
Computer Science and Engineering
Doctor of Philosophy
Document Type:
Date of Defense:
February 21, 2019
Committee Members:
  • Kamesh Madduri, Dissertation Advisor
  • Kamesh Madduri, Committee Chair
  • Jesse Louis Barlow, Committee Member
  • Wang-Chien Lee, Committee Member
  • Necdet S Aybat, Outside Member
  • Kesheng Wu, Special Member
  • Xiaoye Sherry Li, Special Member
  • Gabriel Gomes, Special Member
  • machine learning
  • time series prediction
  • optimization
  • intelligent transportation
  • structural learning
Transportation is arguably speaking one of the most critical functions of human society. It has been an important societal problem since the ancient age, yet the solution is still far from perfect in the twenty-first century. The needs for efficient and safe transportation are ever-growing, due to prolonging life expectancy and diminishing reserves of fossil fuels which most transportation modes rely on in the present day. At the same time, we are facing unprecedented growth of data. Can the society utilize data, a cyber-resource, to solve the physical challenges in modern transportation needs? This question motivates the research in my dissertation. Machine learning, broadly speaking, are algorithms that aim to generalize a set of rules from existent data for describing the data generating process, predicting future events, and producing informed decision making. This dissertation studies previous machine learning methods, improves upon them, and develops new algorithms to contribute in essential aspects of transportation systems. Two important topics in transportation systems are addressed in this dissertation, traffic flow prediction and traffic safety analysis. Traffic flow prediction is a fundamental component in an intelligent transportation system. Accurate traffic predictions are building blocks to achieve efficient routing, smart city planing, reduced energy consumption and among others. Traffic flows are multi-modal and possibly non-stationary due to unusual events. Hence, the learning algorithms for traffic flow prediction need to be robust and adaptive. In addition, the models must be able to learn from latest traffic flow without severely comprising the computational efficiency, in order to meet real-time computation requirements during online deployment. Therefore, learning algorithms for traffic flow prediction developed in this dissertation are designed with the goal to achieve robustness, adaptiveness, and computational efficiency. Traffic safety in transportation systems is as important as efficiency. Rather than predicting the outcome of crashes, it is more valuable to prevent future accidents by learning from past experiences. The second theme in this dissertation studies machine learning models for analyzing factors contributing to the outcome of crashes. The same accident factor may have diverse degrees of influence on different people, due to the unobserved individual heterogeneity. Capturing heterogeneous effect is difficult in general. A viable approach is to impose structure on the unobserved heterogeneity of different individuals. Under some structural assumptions, it is possible to account for the individual differences with respect to accident factors. Temporal learning addressed problems arisen from traffic flow prediction. Structural learning is an approach for modeling individual heterogeneity, aiming to quantify the influence of accident factors.