| As the urban population continues to grow,accurate urban geographic information plays an increasingly important role in our daily lives.It not only provides a daily travel reference for residents,but also serves as a basis of urban resource allocation and scheduling.However,traditional geographic information collection relies on human labor or static geographic sensors.Those methods have high costs,limited spatial coverage and information dimensions.Recently,with the popularity of positioning technology and the development of intelligent industry,there are massive trajectory data generated in the city which are cheap to obtain,have high spatial coverage,and contain spatio-temporal information.Those data perceive the information of urban geospatial entities implicitly,and thus,are a new type of data source for collecting geographic information.However,due to issues like positioning noises,annotation errors,and spatially skewed data distribution,it is a great challenge to collect the geographic information based on the trajectory data.Therefore,we aim to study trajectory data mining techniques for geographic information extraction,and to design methods of geospatial entity discovery and understanding driven by trajectory data,deriving both location and attribute information of urban geospatial entities,which broadly facilitates urban applications,e.g.,route planning,arrival time estimation and urban resource allocation.The detailed research work and contributions of this dissertation contain following four parts.From the aspect of trajectory data-driven geospatial entity discovery,we first study the trajectory data-driven road network generation problem,which is the key technique for road network completion and updates.The existing work is mainly based on the unsupervised learning methods e.g.,clustering,and has difficulties in handling positioning errors and low sampling rate issue.Therefore,we propose a deep learning-based trajectory data-driven road network generation framework that automatically learns the spatial mapping from trajectories to the road network.It consists of two phases: geometry translation and topology construction.In the geometry translation phase,the locations of the road centerlines are predicted by a deep convolutional neural network with the multi-task learning strategy.In the topology construction phase,a trajectory-based post-processing algorithm is used to refine the topological connectivity of the generated road network.The proposed framework has32.3% and 6.5% improvements in the high-sampling-rate and low-sampling-rate trajectory datasets,respectively,and has great value for completing missing roads in residential areas.Correspondingly,we also study the trajectory data-driven location inference problem.The accuracy of commercial toponym resolution services is gradually incapable of the fine-grained calculation in urban intelligent applications.Researchers turned to explore the possibility to infer locations based on couriers’ annotated data in the delivery.However,the effectiveness of existing methods is sensitive to couriers’ annotation behaviors.Therefore,we propose a location inference framework based on couriers’ trajectories and waybills,which is more robust against mis-annotations.It consists of two phases: location candidate generation and location selection modeling.In the location candidate generation phase,a pool of location candidates is constructed by clustering trajectory stay points,and then for each place name,its location candidates are filtered.In the location selection modeling phase,a deep neural network based on an attention mechanism is used to infer the location for each place name.The error of the proposed framework is reduced by 32.7%-47.0% against the best baseline.From the aspect of trajectory data-driven geospatial entity understanding,we take the delivery location understanding as a practice,and first study the trajectory stay point-location association problem,which is a prerequisite for understanding locations using trajectory data.Due to the delivery dynamics,associating stay points to locations essentially is to associate stay points with waybills at those locations,which cannot be achieved based on couriers’ mis-annotations.Therefore,we propose a delivery location-based stay points-waybills association method.The proposed method constructs multiple delivery tasks by grouping waybills in a trip belonging to the same delivery locations and uses a deep learning model to select a stay point near the delivery location of a certain delivery task for association.The model learns the stay reasons based on waybills in the delivery task,uses a spatial grid to encode the distribution of waybills in the neighborhood,and employs a recurrent neural network to capture the temporal dependency of candidate stay points.The association accuracy of the proposed method is 8.5% higher than the best baseline.After that,we take the association results between stay points and delivery tasks as historical observations,and further study the trajectory data-driven service time prediction problem,the accuracy of which directly affects the effectiveness of downstream applications,e.g.,delivery time prediction and parcel allocation.Existing work failed to consider the imbalanced data distribution among locations and the complex delivery circumstances.Therefore,we propose a service time prediction model based on deep meta learning,which predicts the service time at different locations accurately.The proposed model treats the prediction of each location as an independent learning task to understand the heterogeneity of different locations,introduces floor-distribution aware delivery task representation layer to encode complex circumstances,and leverages a location prior knowledge enhanced meta-learning model to learn the shared delivery pattern among locations.The prediction error of the proposed model is reduced by 7.6%-9.5% against the best baseline.It reduces the error of delivery trip time estimation by 14 minutes,which greatly balances the parcel allocation among couriers. |
PDF Full Text Request