Urban events are the sets of real-world occurrences in urban space associated with specific topics. Discovering, understanding, and forecasting various urban events are essential to facilitate smart-city applications and benefit urban life. In recent years, mobile devices and sensors widely located in cities have collected large amounts of data produced in urban space, which form a large-scale, cross-domain, and multi-view data ecosystem. The collected data, termed urban big data, provide an unprecedented opportunity to discover and analyze urban events. In this thesis, we develop data-driven methodologies for urban events mining. We present four of our works on urban anomaly detection, urban location representation learning, road network traffic prediction, and urban events prediction,...[ Read more ]

Urban events are the sets of real-world occurrences in urban space associated with specific topics. Discovering, understanding, and forecasting various urban events are essential to facilitate smart-city applications and benefit urban life. In recent years, mobile devices and sensors widely located in cities have collected large amounts of data produced in urban space, which form a large-scale, cross-domain, and multi-view data ecosystem. The collected data, termed urban big data, provide an unprecedented opportunity to discover and analyze urban events. In this thesis, we develop data-driven methodologies for urban events mining. We present four of our works on urban anomaly detection, urban location representation learning, road network traffic prediction, and urban events prediction, respectively.

First, we introduce an urban dynamic decomposition method to detect urban anomalies. We propose a neural network model to estimate normal urban dynamics using spatiotemporal features and detect abnormal events via residual analysis. We validate the effectiveness of the proposed method with both synthetic and real-world event datasets. Regions are the primary spatial units for urban event analysis. The second work focuses on learning an embedding space from urban big data for urban regions. We propose to construct multi-view relations among urban regions and learn urban region embeddings through a multi-view joint learning model. Experiments on real-world datasets prove the effectiveness of the learned embeddings on downstream applications, including predicting long-term regional event statistics such as crime rates. Traffic prediction is a crucial task for congestion events diagnosis and control. In the third work, we propose a traffic prediction model called Adaptive Spatiotemporal Convolution Network, which models complicated spatiotemporal traffic dependencies adaptively with awareness of the real-time traffic conditions. Experiments on real-world traffic datasets from California and Beijing demonstrate that the proposed model outperforms the state-of-the-art. In the fourth work, we target at location and arrival time prediction of individual urban anomalous events. We design a message passing based mechanism to model the spatiotemporal impacts of historical events. We make joint predictions of event locations and times based on regional states and environmental factors. Our method achieves superior performances on two real-world urban events datasets compared with existing spatiotemporal prediction methods.

In the end, we conclude this thesis with future challenges and research directions related to data-driven urban events analysis.