Forecasting Regional Multimodal Transportation Demand with Graph Neural Networks: An Open Dataset

Nowadays, with the rapid development of the Internet of Things technologies, big spatio-temporal data are able to be obtained from everywhere in our society. Based on such kind of data, transportation demand prediction has drawn increasing attention from the industry for its ubiquitous real-life applications such as traffic scheduling, crowd management and public safety. In this study, we first generate a multimodal transportation demand dataset by using the bike and taxi trip data from New York City. To model the regional transportation demands in non-Euclidean space, we employ Graph Neural Networks including STGCN, DCRNN, and Graph WaveNet and further implement several generic performance-boosting strategies by respectively utilizing multi-source (bike-inflow, bike-outflow, taxi-inflow, taxi-outflow), multi-graph (adjacency matrix and origin-destination matrix), and meta-information (dayofweek, hourofday, isholiday). The open dataset and the deep models implemented with PyTorch are now available on https://github.com/Evens1sen/Deep-NYC-Taxi-Bike.