Spatio-temporal problems arise in broad areas of environmental and transportation systems. These problems are challenging, because of both spatial and temporal neighborhood similarities and correlations. We consider traffic data, which is a complex example of spatio-temporal data. Traffic data is geo-referenced time series data, where fixed locations have observations for a period of time. Traffic data analysis and related machine learning tasks have an important role in intelligent transportation systems, such as designing navigation systems, traffic management, control systems and in the future will be essential for setting appropriate anticipatory tolls. Recent data collection methodologies dramatically increase the volume of available spatio-temporal data, which require scalable machine learning models. Moreover, deep learning models outperform traditional machine learning and statistical models due to their strong feature learning abilities in spatial and temporal domains. Increases in available data and recent advances in deep learning models in spatio-temporal domains are the main motivations of this dissertation.
We first study, non data-driven and optimization-based solutions for the network flow problem, which appears in a wide range of applications including transportation systems and electricity networks. In these applications, the underlying physical layer of the systems can generate a very large graph resulting in an optimization problem with a large decision variable space. We present a distributed solution for the network flow problem. The model uses cycle basis and an alternating direction method of multipliers (ADMM) method to find a lower computational time and number of communications, while obtaining a centrally optimal solution.
Second, we attempt to obtain spatio-temporal clusters in traffic data, which represent similar traffic data in terms of both spatial and temporal similarities. Clustering of traffic data are used to analyze traffic congestion propagation and detection. We obtain spatio-temporal clusters using a modification to Deep Embedded Clustering, which considers both spatial and temporal similarities in latent features. Also we define new evaluation metrics to evaluate spatio-temporal clusters of traffic flow data.
Third, when sensors collect spatio-temporal data in a large geographical area, the existence of missing data cannot be escaped, which negatively impacts of prediction models. Here, we investigate the problem of incorporating both spatial and temporal contexts in missing traffic data imputation using convolutional and recurrent neural networks. We propose a convolutional-recurrent autoencoder for missing data imputation, and illustrate the performance of autoencoders for missing data imputation in spatio-temporal data.
Finally, traffic flow prediction has an important role in diverse intelligent transportation systems and navigational systems. There is a large literature on this problem. However, the problem is challenging for high-dimensional traffic data. We explicitly design the neural network architecture for capturing various types of spatial and temporal patterns. We also define evaluation metrics for spatio-temporal forecasting problems to better evaluate generalization of the model over various spatial and temporal features.