| With the development of urbanization,the central clustering effect of large cities becomes more and more obvious.As cities develop rapidly,problems such as traffic jams and delays become obstacles for cities,which pose new challenges to the planning and management of cities.Many of the sensors deployed in modern cities,such as mobile phones,GPS devices,surveillance cameras,etc.,generate large amounts of data every day.These data record the operation and changes of the city.In recent years,the big data industry chain has become more mature.Analyzing and managing cities by utilizing various big data in cities has become a hot research topic.Urban functional regions mining is one of the important applications.The discovering of urban functional regions aims at the use of city-neutral location-related data,such as points of interest,human trajectories,and vehicle trajectories,to classify regions with similar functions.At present,the mainstream approach is to treat the region as a document and the function of the region as the topic of the document.The various types of data in the region can be regarded as the words or metadata of the document,and topic models in natural language processing can be applied to achieve the discovering of functional areas.However,the existing methods require feature engineering by hand,which will construct features of limited expressiveness,and apply probabilistic topic models such as latent Dirichlet allocation.The probabilistic topic model is based on the bag-of-words hypothesis,ignoring the contextual relationship of the words,while the trajectory features is a time series in the regional document model,and there is a potential link between its contextual semantics and the topic of the region.Therefore,the use of a probabilistic topic model for functional area division has limitations.This thesis focuses on the shortcomings of the existing methods and proposes improvements.The main research work is as follows:(1)In order to overcome the limited expression capabilities of artificially constructed features,a deep learning framework was introduced to automatically build powerful features and extract more abundant information from the original datasets.For the disadvantage that the existing model cannot extract the context information of trajectories,we proposed a structure using a deep neural network based topic model to infer the topics of city regions,so as to utilize the potential context semantics in the trajectories.(2)Aiming at the defects of deep neural network base topic models such as expensive computation,oblivion when process long sequences and unable to fuse multiple datasets,we proposed a more advanced framework,Mobility Pattern Embedded Topic Model(MPETM),that combined deep learning and probabilistic topic models.The word embedding algorithm in deep learning was used to embed trajectories,and then used the probabilistic topic model to fuse the embedded trajectories and multi-source dataset to excavate urban zones of different functions.(3)Generalize the MPETM framework.Many applications in the field of urban computing or data mining involved the integration of time series data and other data.Through the MPETM framework,the potential contextual information and global semantics in the data can be effectively extracted.In this paper,experiments were performed on large-scale datasets.By comparison with the clustering algorithm,Dirichlet multimodal regression topic model and deep learning based topic model,the MPETM framework has superior accuracy. |
