Graph deep learning models for network-based spatio-temporal data forecasting : from dense to sparse

Ongoing rapid urbanisation has modernised people’s lives but also engendered the increase in traffic congestion, energy consumption, air pollution, urban crimes, road incidents, etc. With the advance in the Internet of Things (IoT) and 5G, massive geotagged and timestamped (spatio-temporal) data have been collected to monitoring urban environment and processes. There are increased interests in developing urban spatio-temporal (ST) forecasting to make cities greener, safer, and smarter. Recently, deep learning (DL) has been used widely in urban environmental monitoring because of its powerful capability in modelling complex and high dimensional data. Its full potential for urban process prediction is yet to develop due to the irregular network-based spatial structure in many urban processes, temporal non-stationarity, ST heterogeneity and data density variation. In addition, real-world applications at city-scale require fast (or near real-time) training and prediction, capable of dealing with abnormal conditions in real-world scenarios (e.g. missing data and non-recurrent events). To address the challenges, this thesis has developed cutting-edge graph DL models to forecast large-scale urban processes on networks. The contributions of this study are summarised from two aspects. First, from a methodological perspective, we use graphs to unify the representation of all ST urban processes, either dense or sparse. A number of novel contributions are then made towards DL technique by expanding and adapting DL to a spatiotemporal framework for different types of network-based ST forecasting tasks. We propose unified DL models with novel spatial or spectral graph convolutions to forecast both directed and undirected dense urban processes on networks, addressing issues including non-stationary temporal dependency modelling, network-structured spatial dependency modelling and ST heterogeneity. We further tackle the data sparsity issue by developing the first graph DL model with an innovative localised weight sharing graph convolution. The proposed models have scalable structures that can produce citywide ST forecasts in a timely and accurate fashion. The modularity of these models allows to deal with missing data and incorporate external factors for robust forecasting under abnormal conditions, which enhances the chance of the models being used in real-time urban applications. From the urban application point of view, various large, real-world urban datasets, including traffic and crime cases, have been employed to validate that the proposed models can outperform various state-of-the-art benchmarks in terms of accuracy and efficiency. The results derived through these forecasting techniques can be used to address many key growth areas in urbanisation, like human mobility, transportation, and public safety, which has the potential to facilitate future policies and improve the well-being of societies.