Extracting new urban patterns in cities: Analysis, models and applications

Abstract

Smart city initiatives rely on real-time measurements and data collected by a large num- ber of heterogenous physical sensors deployed throughout a city. The data gathered by physical sensors can capably identify important events in cities, but seldom explain the underlying reasons behind such events. In other words, physical sensors can identify what happens, but may be unable to explain why or how specific events occur or patterns emerge. The rise of Location-based Social Networks (LBSNs) has allowed millions of dwellers and visitors of a city to share their observations, thoughts, feelings, and experi- ences, or in other words, their perceptions about their city through social media updates. LBSNs data represents a treasure which is still under explored, especially as the added location dimension on social networks bridges the gap between the physical world and the digital online social network services, potentially leading to the emergence of new types of applications. This thesis shows how the use of this powerful LBSNs data coupled with machine learning techniques, can lead to the extraction of new urban patterns in cities. The thesis starts by leveraging the power of Deep Learning and in particular, Deep Belief Networks, for extracting a new urban pattern which is called Socio-demographic Regional Patterns. It is shown for the first time that it is possible to extract a unique pattern for various regions in cities of very close spatial proximity. The five boroughs in New York City are considered in a case study with an emphasis on extracting a unique pattern for each of the boroughs. Second, a new approach is introduced that discovers functional regions that not only change across space but time as well. It is shown with the proposed approach that it is possible to extract different functionalities for the same physical regions during the day. This type of new urban pattern is called, Temporal Functional Regions Patterns. Next, a new approach for Recognizing Recurrent Crowd Mobility Patterns in cities is introduced for illustrating how crowd shifts across space and time with various crowd level intensities. In addition, it is shown that the correlation between the extracted crowd mobility patterns and the temporal functional regions patterns provides further new insights into the motivation behind crowd mobility. Finally, it is shown how some of these extracted patterns can brought together to solve a domain specific challenge (Network Demand Prediction). To do so, a new deep learning based-approach (titled ST- DenNetFus) is introduced for fusing some of the extracted urban patterns with network demand data achieving a higher level of accuracy on predicting the network demand across cities compared to without fusing these patterns.