Graph Neural Networks for Transit Passenger Flow Prediction

点击翻译按钮获取中文摘要

The transportation sector is the second largest contributor to GHG emissions in Canada. Public Transportation is key to Canada's goals to reduce GHG emissions in the future. Improving transit services can thus help attract people towards public transportation thereby contributing to transportation sector emissions reductions. A key contributing factor to improving transit operations is the reduction of crowding on buses. Central to being able to do this is being able to accurately predict passenger levels on buses so that transit operators can adjust their operations to manage and reduce crowding. At the same time, the past twenty years has seen extraordinary advances in the development and use of Artificial Intelligence and "deep learning" techniques. Many of the advances from these techniques are thanks to the use of convolutional neural networks (CNNs) in image processing and classification. These methods are well adapted to situations in which data can be represented in regular (e.g. grids) Euclidean space, but are less well adapted to situations (like transportation) in which data is more appropriately described as networks or "graphs" that include links (road segments) and nodes (intersections). Luckily, advances in deep learning based on CNNs have spurred innovations leading to the ability to use deep learning techniques with graphs and graph data. This has been done through advances in graph representation learning. Graph representation learning has enabled the power of convolutional neural networks to be applied to graph learning and thus to allow such deep learning techniques called Graph Neural Networks (or GNNs) to be applied to network related applications such as those in transportation. While there has been an explosion in the use of deep graph-based approaches in road performance prediction, transit has received less attention, and bus transit has been almost unexplored. The goal of this research program is to extend the use of GNNs in public transportation and particularly bus transportation to predict bus passenger levels using both historical and real-time Big Data in the form of automatic passenger counts and automatic vehicle location. This will help bring the power of deep learning to public transportation planning, make contributions in graph representation learning in transportation and computer science more broadly, and help Canada meet its GHG emissions goals.