Robust and efficient spatio-temporal graph learning

Abstract

The rapid growth of urban data has sparked significant interest in the field of urban studies. As cities consist of diverse regions such as business districts, residential areas, and more, generating high-quality region spatial-temporal graph embeddings is crucial for understanding the underlying structures of urban environments. These embeddings have the potential to contribute to the development of smarter and more sustainable cities. Additionally, they directly impact various downstream prediction tasks, including traffic flow prediction, crime prediction, and others. The proliferation of mobile computing technologies has resulted in an unprecedented availability of urban data, such as taxi trajectories and Point-of-Interests (POIs), providing valuable support for exploring and analyzing region embeddings. However, generating high-quality region embeddings from abundant city data presents challenges, such as data sparsity. In this thesis, we propose three novel frameworks for robust and efficient spatio-temporal graph learning. Our approach utilizes graph neural networks (GNNs) to capture spatial dependencies and temporal dynamics in the graph data. We incorporate robust regularization techniques to enhance the model's resilience to noise and outliers. Additionally, we introduce two new frameworks for anomalous subtrajectory detection and traffic prediction. Through extensive experiments on real-world spatial-temporal datasets, we demonstrate the superior robustness and efficiency of our proposed framework compared to state-of-the-art methods. Our approach not only achieves accurate predictions but also ensures the model's stability and reliability, even in the presence of noisy or outlier data points. Furthermore, our efficient graph sampling strategy enables faster training and inference, making our framework applicable to large-scale spatio-temporal graph datasets. The robustness and efficiency of our proposed framework make it well-suited for various applications, including urban planning, transportation management, and environmental monitoring. By accurately capturing the complex relationships and dynamics in spatio-temporal data, our framework provides valuable insights and supports data-driven decision-making processes.