Learning On Spatial-Temporal Data For Prediction

Spatial-temporal data prediction is one of the core research issues in the field of spatial-temporal data mining,which can find numerous application scenarios in many domains such as transportation,meteorology,and medicine.Especially,in the transportation field,designing suitable learning models for different types of spatial-temporal data to serve various prediction tasks is valuable.Because such study can accelerate the realization of intellectualization of the transportation information systems,so as to enhance the safety and operation efficiency of the systems,and improve the experience of traffic participants.In the process of advancing the development of modern intelligent transportation systems,how to achieve accurate spatial-temporal data prediction to meet the increasing various prediction demands and high-precision prediction requirements is a crucial problem that needs to be solved urgently.Traditional data modeling and prediction methods fail to effectively learn the important patterns of various spatial-temporal data types,such as spatial-temporal correlation,periodicity,heterogeneity,etc.,resulting in unsatisfactory prediction results,so they cannot meet the actual application requirements.In this dissertation,a series of novel spatial-temporal deep-learning based prediction models are proposed according to the characteristics of various spatial-temporal data types,including the spatial-temporal raster data,spatial-temporal grid data and spatialtemporal event data,to meet the prediction requirements of real-world applications in the transportation field.The main contributions of this paper are summarized as follows:1.For the spatial-temporal raster data prediction,in order to capture the correlation and heterogeneity of spatial-temporal raster data in both space and time,a novel end-to-end deep learning based model,called ST-3DNet are proposed.ST-3DNet introduces 3D convolutions to automatically capture the correlations of traffic data in both spatial and temporal dimensions.And a novel recalibration(Rc)block is proposed to explicitly quantify the different contributions of the correlations in space.Considering the two temporal properties of spatial-temporal raster data,i.e.,local patterns and periodicity patterns,ST-3DNet employs two components consisting of 3D convolutions and Rc blocks to respectively model the two properties.The experiments on three real-world traffic datasets demonstrate that ST-3DNet can improve the prediction accuracy of spatial-temporal raster data.2.For spatial-temporal graph data prediction,in order to model the dynamics of spatial-temporal graph data along both temporal and spatial dimensions,and to capture the periodicity and the spatial heterogeneity,an Attention based SpatialTemporal Graph Neural Network(ASTGNN)is proposed.Specifically,in the temporal dimension,a novel self-attention mechanism is designed to utilize the local context,which is specialized for numerical sequence representation transformation and enables the prediction model to capture the temporal dynamics of any two data points.In the spatial dimension,a dynamic graph convolution module is developed,which employs self-attention to capture the spatial dynamics.Furthermore,temporal and spatial embedding modules are introduced to explicitly model the periodicity and capture the spatial heterogeneity.The experimental results on five real-world traffic flow datasets demonstrate that ASTGNN outperforms the stateof-the-art baselines.3.For spatial-temporal event location prediction,in order to comprehensively capture the spatial-temporal context and the spatial-temporal correlations of asynchronous mobility events,and to effectively learn users' mobility regularities from the mobility behaviors with high diversity and variation,a deep generative network(LDGN)is proposed,which can predict the location of next mobility event given the historical mobility events.Specifically,a novel spatial-temporal heterogeneous graph embedding module is designed to provide a comprehensive description of the context where mobility event happens.A hierarchical spatial-aware transformer accompanied with a continuous time embedding is designed to encode the asynchronous and spatial-temporal correlated historical mobility events.Besides,LDGN explains that the generation of mobility events is driven by users' requirements,and introduces a hidden variable topic subject to the gaussian mixture distribution to reflect users' different mobility requirements.The experimental results on three real-world datasets demonstrate the superiority of LDGN.4.For spatial-temporal event time prediction,based on the study in Chapter 5 which has already modeled the spatial-temporal context,captured the spatial-temporal correlations of asynchronous mobility events,and considered the high diversity and variation of user mobility behaviors,a deep generative network(TDGN)is further proposed,which can predict the happening time of next mobility event given the historical mobility events.TDGN is designed based on the architecture of LDGN,to take advantage of LDGN on effectively modeling the spatial-temporal mobility event.TDGN further employs a mixture log-normal distribution to model the time intervals between successive mobility events,which is able to capture users' complicated mobility regularities in the temporal dimension.The experimental results on three real-world datasets demonstrate the superiority of TDGN in learning the temporal patterns of user mobility behavior and making accurate event time predictions.