Algorithms for Context-Sensitive Prediction, Optimization and Anomaly Detection in Urban Mobility
Transportation infrastructure is a complex human cyber-physical system that is currently facing significant challenges in many communities around the world. The problem emanates from increased congestion, which results in large-scale inefficiencies, including significant personal, health and environmental costs. The human integrated nature of this resource constrained system allows communities to go beyond the traditional mechanisms of adding infrastructure which is often expensive and difficult to build and embrace data-driven smart solutions that focuses on providing a robust decision support system, which can enable humans to use and optimize the system more efficiently. However, there are several challenges that arise due to the heterogeneity, sparsity, and noise in the data collected in an urban environment. This dissertation examines a unique application platform called transit-hub that enables (1) integration of spatially and temporally distributed sensor streams, (2) integration of simulation-based decision support systems, and (3) development of experiments to understand how advanced decision support tools improve the utilization of the transportation infrastructure. We designed data mining and machine learning techniques for context-sensitive prediction of long-term, short-term and real-time delays in sparse public transit networks. In order to solve data sparsity issues, shared route segment networks and multi-task neural networks were developed. Further, we integrated algorithms for analyzing the performance of public transit networks and developed mechanisms to optimize the on-time performance under uncertainty of traffic and weather conditions. Heuristic search algorithms as well as sensitivity analyses of the hyper-parameters were also developed. Robust detection of anomalous operations of transit networks over a large metropolitan area was enabled by using deep learning techniques within the platform. A specific innovation is to set up the problem of identifying the non-recurring traffic congestion as an image classification task and to use convolutional neural networks to explain the congestion.