Urban Region Representation Learning Using Human Trajectories Of Multiple Granularities:Methods And Applications

Humans have accumulated abundant behavioral trajectory data in urban economic life with the popularization and implementation of various big data intelligent applications.Human trajectory data reveals the socio-economic attributes and dynamic evolution of urban regions.Extracting feature information from human trajectory data through urban region characterization deeply portrays and understands urban regions,provides a basis for urban design and planning,and helps assist urban managers’ decision-making.Due to the various data requirements and privacy protection issues,the human trajectories exhibit multi-grained characteristics.There are two kinds of granularity trajectory data that are mainly used in research and practice:fine-grained human trajectory data,which describes the detailed mobile trajectory of a single subject over a period,such as navigation data and cell tower data;coarsegrained human trajectory data,due to the constraints of privacy protection and other factors,many data only retain the starting and ending information of travel,but ignore the intermediate trip information,such as taxi trip data which only contains the information of passengers pick up and drop off locations information.In view of a series of challenges brought by the multi-granularity characteristic of human trajectory data,(1)this paper proposes a general framework for urban region representation learning using multi-granularity human trajectory data,which constructs multi-view for regions based on trajectory data to reveal regional attributes from multiple perspectives and achieves information complementarity and enhancement between multi-view through contrastive learning.To verify the effectiveness of the framework,(2)this paper refines components of the framework and proposes a multi-view representation learning model for fine-grained human trajectory data,which constructs a transition view,a temporal view,and a spatial view for regions,and uses a contrastive learning method based on graph structure enhancement and mutual information maximization theory to learn multi-hop dependency relationships and spatio-temporal correlations between regions;(3)meanwhile,this paper refines components of the framework and proposes a multi-view representation learning model for coarse-grained human trajectory data,which constructs population inflow view,outflow view,and spatial view for regions,and calculates the losses intra and inter view through a dual multiple loss function based on triplet losses to capture co-occurrence relationships between regions.This paper conducts experimental verification and application on cell tower data and taxi trip data in two real cities,Shenzhen,Guangdong Province,China,and Manhattan,New York City,United States.The performance of models has been analyzed through multiple urban downstream tasks(e.g.land use classification,population density estimation,and crime prediction),and the experiment results show that the models outperform the popular baseline method.Two urban computing applications,region similarity search,and urban functional area identification are applied to verify the urban region representation learned in this paper is consistent with the long-term planning and development goals of the city and meets the practical management needs of the government.