Multi-view Joint Learning Method For Region Representation Based On Graph Structure

Regions are essential elements of cities,where people work,live and play.Studying the learning problem of region representation can better understand the patterns,structures and dynamics of cities,support urban planning,and make cities smarter and sustainable.In recent years,more and more urban data(e.g.,vehicle trajectories,points of interest,and check-in records)are rapidly accumulating in different forms as mobile sensing technologies become more popular.The large amount of multimodal data brings new challenges to study area representation,attracting researchers to develop advanced techniques to better capture various area correlations.Existing work is mostly based on single-view data to mine the inherent attributes of regions or correlations among regions,and the performance of the models is largely constrained;even using multi-view data,existing work still suffers from many shortcomings:it only focuses on attribute information such as points of interest in regions,ignoring the phenomenon of region spatial imbalance;it does not consider the differences in human mobility in terms of temporal dynamics.In addition,existing work does not pay attention to the cooperation between views,and the information on the intra-region attributes and inter-region correlations is not deeply integrated due to the lack of information sharing and dissemination.To address the problems of existing work,this thesis focuses on understanding and mining multimodal data relationships in cities,and innovatively proposes a multi-view joint learning model to learn embedding representations of urban regions,with the following main work.(1)This thesis constructs multiple views from two aspects,region attributes and human mobility,to guide region representation learning.In terms of region attributes,this thesis proposes a region static feature description method,which eliminates the region spatial imbalance and describes the attribute features of regions more accurately.Meanwhile,interregion distance similarity is used to enhance the attribute feature representation of regions.In terms of human mobility,the flow distribution vector is constructed to describe the inflow and outflow patterns of regions within a fixed period of time,which is more accurate compared with the traditional approach.(2)Graph structures are used to represent attributes to better capture the relationships between underlying regions,and a graph attention mechanism is applied to learn region representations in each correlation view constructed.In addition,this thesis also designs a joint learning module to facilitate the learning of individual views through a cross-view information sharing layer,and a fusion layer to effectively combine multiple views to finally obtain a unified region representation.(3)The performance of the model in downstream tasks such as land use classification,region popularity prediction and crime prediction is evaluated using real datasets.Extensive experimental results show that the performance of the proposed model is significantly improved on all three tasks compared to the state-of-the-art approaches.