The availability of rich information about fine-grained user mobility in urban environments from increasingly geographically-aware social networking services and the rapid development of machine learning applications greatly facilitate the investigation of urban issues. In this setting, urban computing emerges intending to tackle a variety of challenges faced by cities nowadays and to offer promising approaches to improving our living environment. Leveraging massive amounts of data from geo-social networks with unprecedented richness, we show how to devise novel algorithmic techniques to reveal underlying urban mobility patterns for better policy-making and more efficient mobile applications in this dissertation.

Building upon the foundation of existing research efforts in urban computing field and basic machine learning techniques, in this dissertation, we propose a general framework of urban computing with geo-social network data and develop novel algorithms tailored for three urban computing tasks. We begin by exploring how the transition data recording human movements between urban venues from geo-social networks can be aggregated and utilised to detect spatio-temporal changes of local graphs in urban areas. We further explore how this can be used as a proxy to track and predict socio-economic deprivation changes as government financial effort is put in developing areas by supervised machine learning methods. We then study how to extract latent patterns from collective user-venue interactions with the help of a spatio-temporal aware topic modeling approach for the benefit of urban infrastructure planning. After that, we propose a model to detect the gap between user-side demand and venue-side supply levels for certain types of services in urban environments to suggest further policymaking and investment optimisation. Finally, we address a mobility prediction task, the application aim of which is to recommend new places to explore in the city for mobile users. To this end, we develop a deep learning framework that integrates memory network and topic modeling techniques. Extensive experiments indicate that the proposed architecture can enhance the prediction performance in various recommendation scenarios with high interpretability.

All in all, the insights drawn and the techniques developed in this dissertation make a substantial step in addressing issues in cities and open the door to future possibilities in the promising urban computing area.