| In the context of continuous urbanization and China’s vigorous development of urban digital economy,in-depth mining and analysis of urban spatial-temporal big data has become a crucial engine to drive social,economic,and technological progress.In recent years,through mobile sensing,the Internet,and various fixed and mobile sensing devices deployed in cities,massive urban spatial-temporal data have been collected and utilized to reflect the dynamics of urban development in different fields.Through modeling and prediction of massive urban spatial-temporal data,many research efforts have yielded important values in the fields of transportation,communication,environment,economy,and resource optimization,significantly improving the efficiency of urban operations and the quality of life of residents; at the same time,considering the huge scale of cities,the transfer and large-scale deployment of prediction models are also important.Therefore,this thesis focuses on the research for urban spatial-temporal data prediction and knowledge transfer in different scenarios.The main challenges of this research topic are as follows: Firstly,urban spatial-temporal data are of high dimensionality and large scale,and there are multi-scale correlations,which make it difficult for existing models to be directly applicable.Secondly,the spatial-temporal data acquisition and model deployment in new scenarios are costly,and there is a lack of efficient urban spatial-temporal knowledge transfer methods.Finally,considering the complex and diverse urban dynamics,which are closely coupled with multi-modal information such as weather,environment,regional functions,and government policies,it is very difficult to fuse the spatial-temporal features of multi-modal data.To address these challenges,this thesis investigates the prediction and knowledge transfer problems of spatial static-temporal dynamic data(a.k.a.multivariate time series)and spatial dynamic-temporal dynamic data(a.k.a.spatialtemporal dynamic data)in cities,respectively,and conducts experimental validation under specific application scenarios.The main contributions of this thesis are summarized as follows.1.The sources,types,characteristics,and typical application scenarios of urban spatialtemporal data are outlined,existing work is introduced,and the pressing issues and challenges are described.2.The problem of multivariate time series correlation modeling is investigated.For the radio spectrum prediction scenario,a neural network based on a two-stage attention mechanism is proposed in this chapter.The model takes into account the complex temporal-frequency domain correlation in radio spectrum signals,and introduces two attention mechanism modules to model the signal correlation in frequency and temporal domains,respectively.In addition,an image-processing-based signal detection module is designed to quickly locate the signals from the raw radio spectrum data for subsequent modeling and analysis.Experiments are conducted with real-world data to verify that the model achieves about6%～17% improvement in prediction accuracy over the baseline algorithm for the radio spectrum prediction task,and finally the interpretability of the model is analyzed in conjunction with a visualization approach.3.A instance-based framework for spatial-temporal knowledge transfer is studied oriented to radio spectrum prediction scenarios.The knowledge transfer problem between multiple sites and multiple frequency bands radio spectrum is further considered based on single-site multivariate time series modeling.A weighted transfer learning framework is proposed,in which data from different domains are weighted according to their similarity at the data level,so that the instances from the source domain containing more common features with the target domain are strengthened.Experimental results show that the proposed framework can be robustly applied to various scenarios in the real world and improve the performance in the target domain.4.The problem of multi-scale modeling of spatial-temporal dynamic data is investigated for the radio map reconstruction scenario.I design a neural network based on spatialtemporal residual module and image super-resolution methods,which can effectively extract valuable features from coarse-grained spatial-temporal dynamic data and reconstruct the fine-grained radio maps.Experimental results show that our approach outperforms the baseline model in multiple task settings with about 6%～19% improvement in reconstruction accuracy.Besides,I conduct a case study that illustrates that the model effectively improves the accuracy of the Wi-Fi fingerprint-based localization system with about 5m localization accuracy.5.The problem of feature-based spatial-temporal knowledge transfer is further investigated under the scenario of radio map reconstruction.Considering the difficulty of collecting fine-grained radio map data with labels in real systems,I introduce an unsupervised domain adaptation approach to increase the model’s generalization capability in the absence of any labels from a new scenario.The method employs an adversarial learning framework for training,so that the data in the source and target domains are pairwise in the feature space extracted by the encoder,based on which the model can then efficiently generalize from the source domain to the target domain,saving the consumption of model training and data collection.6.The problem of fusing spatial-temporal features of multi-modal data is studied under the urban crowd flow prediction scenario during COVID-19.I design a transformer-based generative adversarial network while adding a gated fusion module to model the impact of multi-source heterogeneous information(e.g.,government policies,COVID-19 conditions)on crowd flow,which can accurately predict the pattern of crowd flow under different policies in the next wave of the epidemic.The model also allows policy-makers to better assess the potential impact of various policies on crowd flow.More appropriate policies can be developed to control the propagation of the epidemic while mitigating the effects on people’s normal lives. |
