Research On Key Technologies Of Spatio-temporal Data Mining For Urban Computing

Urban computing is an emerging intersection of computer science with urban backgrounds,integration with urban planning,transportation,energy,environment,economics,and sociology.The task of urban computing is to first analyze and process and display big data by sensing and acquiring all kinds of big data generated in the city,and then using efficient data management technology,advanced algorithms and novel visualization techniques to solve problems and challenges in the city,such as traffic congestion,air pollution,poor planning,etc.There are many kinds of big data generated in cities,such as trajectory data generated when people or cars move,news released by major media platforms,and transaction records that are updated in the stock market.We are devoted to the study of one type of city big data-temporal-spatio big data,that is,data containing both time and space information,such as track points,subway card data,and so on.Through the research on key technologies of spatio-temporal big data mining,it provides a series of solutions for many problems in the city.This paper focuses on three issues from the micro to the macro level in the city: continuous proximity detection of mobile users,regional taxi demand forecasting,traffic prediction between origin and destination regions.We develop effective algorithms and frameworks based on the latest technical means,which results in reducing energy consumption,mitigating uneven supply and demand,and assisting urban design.The main work and innovations of Our thesis are as follows:1.The continuous proximity detection problem of mobile users is an important issue for urban computing at the mobile user level.Almost all existing methods adopt the safe region technology.However,safe regions in the existing work have unreasonable shapes or assumptions,resulting in a large communication overhead.Aiming at this problem,this paper proposes a predictive safe region-based proximity detection method called PSR-PD.In PSR-PD,a predictive safe region is designed,which combines the trajectory prediction algorithm to fully capture the user's motion pattern and minimize the probability of users getting out of the security domain,thus greatly reducing communication overhead.The tests on the four real data sets show that the performance of the algorithm framework is superior to the previous STRIPE algorithm,FMD and CMD algorithms.2.The problem of taxi demand forecasting is an important issue for urban computing at the regional level.Existing methods either consider only the effects of temporal relationships or do not consider the effects of non-adjacent regions when capturing spatial dependencies.For the capture problem of mutual influence between nonadjacent spatial regions,this paper first proposes to apply the graph neural network to the forecast of taxi demand,and regard the area in the whole city as the vertices in the map,and creatively propose to use the origin-destination relationships to define edges in the graph,so that non-adjacent regions can pass through the connection of the edges.All this construct the framework OD-TGAT.Large-scale tests on the New York taxi dataset show that the spatial-temporal deep learning model of this paper predicts taxi demand by biggest 9.3% compared to the comparison algorithms such as MLP,XGBoost,ST-Res Net.3.The problem of traffic prediction between the origin and destination regions is an important issue at the regional-regional level of urban computing.Existing work has been studied on regional flow forecasting or origin-destination intersection flow forecasting,but no intra-regional flow forecasting studies have been conducted.Therefore,this paper is the first to propose a origin-destination regional traffic forecasting problem.Aiming at this problem,this paper proposes a multi-view spatial-temporal deep learning framework MV-STAN that differentiates three traffic dependencies.After dividing the city into grids,different maps are generated under different dependencies,and then results are learned through the respective spatial-temporal deep learning networks,and then merged to form the final prediction results.Tests on two large-scale datasets show that the multi-view spatiotemporal learning framework has up to at least 6.8% accuracy improvement for origin-destination region traffic forecasts compared to other recent algorithms such as ST-DNN,LSTM,GRU.