Research On Some Problems Of Urban Computing

An in-depth understanding of the impact of the urban built environment on the urban real estate market and crowd movements will bring benefits to the government in dealing with many urban challenges in terms of urban planning and policy formulation.This thesis aims to study the effect of the urban built environment on the real estate market as well as crowd movements through urban computing methods.The content is as follows:(1)Aiming at the submarket effect brought by the heterogeneity of the built environment to the real estate market,a Bayesian network-based model is proposed,which can simultaneously use multiple submarket conditions to model the submarket so fully capture the submarket effect.At the same time,aiming at the problem of analyzing the relationship between submarkets and the structure of the real estate market,a probabilistic hierarchical clustering algorithm is proposed for the hierarchical structure of the real estate market.Through the Bayesian network-based model proposed in this thesis,multiple submarket conditions are introduced,and the probability dependency in the Bayesian network is used to constrain the latent variables.In order to analyze the structure of the real estate market,the Bayesian hypothesis testing framework is used to step-by-step find the two sub-markets with the most similar submarket conditions and then aggregate them.Experiments in actual data sets show that the Bayesian network-based model proposed is not only advanced in predictive performance,but also it can provide domain experts with useful insights into the real estate market in terms of interpretability.By comparing urban theory and government planning,the effectiveness of the real estate market structure inference algorithm proposed in this thesis and the value of decision-making support by comparing the urban theory and governmental planning respectively are verified.As far as the author of this thesis knows,it is the first time to propose a method to analyze the structure of the real estate market by analyzing the relationship between submarkets.(2)Aiming at the need to analyze the built environment problem through the human mobility pattern in the urban planning process,this thesis proposes a method based on nonnegative matrix decomposition to quickly retrieve the target area by analyzing the population dynamic pattern.At the same time,it proposes a Motif-based characterization of daily individual movement patterns so as to analyze travel characteristics of local areas in the city.Specifically,in terms of population dynamic pattern,this thesis proposes the use of non-negative matrix factorization to extract basic patterns in crowd dynamics.Due to the parts-based representation ability and good interpretability of non-negative matrix factorization,domain experts,such as urban planners,can well understand the basic patterns of crowd dynamics obtained by non-negative matrix factorization,and annotate each basic pattern of crowd dynamics with semantic tags,such as morning peak mode,evening peak mode,etc.The crowd dynamics can be explained as the linear superposition of the basic patterns of crowd dynamics.Based on this,a retrieval method is proposed to linearly combine the basic patterns of specific population dynamics to help domain experts find interesting built environment areas,such as commuter towns.In terms of individual movement pattern mining,this thesis uses motif to characterize individual travel trajectories.Through the stay point detection,the individual’s daily travel stay point sequence is obtained from the mobile phone signaling data,and then the dwelling place in the stay point sequence is identified through the dwelling detection.Finally,a motif with a dwelling point is formed,and the motif is mapped to the urban space according to the residential area,so that the travel characteristics of the local area of the city can be analyzed.(3)Aiming at the problem that the prediction of commuter flow based on the built environment is affected by spatial correlation,this thesis proposes a model based on graph neural network to capture the complex spatial correlation in the built environment.Specifically,a data representation graph structure called geo-adjacency network is constructed.Then,a gated geo-contexture multi-task embedding learner is proposed to learn the embeddings of various blocks in the geo-adjacency network.Due to the way the geo-adjacency network is constructed,the embedding learner can automatically capture complex spatial correlations.Also,a multi-task learning framework is added to enhance the ability of embedded representation to predict commuting flow.Finally,the commuting flow is predicted through a gradient boosting machine based on the learned embeddings.Real-world data sets are used to evaluate the model in two representative cities(i.e.,New York City and Los Angeles).The experimental results show the necessity of modeling spatial correlations and the advancement of the method proposed in this thesis.Finally,this thesis puts forward the future research framework and key issues of urban science,a new interdisciplinary subject.