Research On Spatio-Temporal Correlation Mining In Urban Computing

With the development of modern technology,the amount of sensor data generated in cities is increasing,which contains important information such as resource usage,population density,weather conditions,and traffic status.Analyzing and mining these data can provide powerful support for urban management and planning.To fully utilize these limited urban data,it is necessary to model and learn the correlation between different features,in addition to mining the knowledge carried by individual data.Modeling the correlation structure between different features can better complement and enhance the data.Traditional models increase the dimension of input data by using data concatenation to expand the model’s field of view.Although this method can consider the correlation between multiple features and their impact on the results,it is inefficient and lacks interpretability.Considering the natural structured spatio-temporal characteristics of sensors in cities,how to make full use of domain knowledge to model the correlation structure more accurately and improve the efficiency of data utilization,so that the results are more accurate and have a certain interpretability,is a challenging problem that needs to be overcome.Existing methods are able to mine the information transmission correlations between nodes based on the existing correlation structure to enhance the expressiveness of the model.However,in scenarios where the correlation structure is unknown,it is still difficult to use limited data to compensate for data sparsity in certain spatio-temporal ranges.Furthermore,due to the lack of a universal way of constructing correlations,correlation-based reinforcement learning is also difficult to apply in multi-category feature scenarios.In addition,most methods are unable to dynamically mine complex and variable spatio-temporal patterns due to the lack of capturing differences between spatio-temporal nodes.Furthermore,as urban scenarios become increasingly complex,most algorithms still rely on feature extraction and concatenation from different data sources,lacking the integration of multiple data sources with real-world scenario-related structures.To address the issues in spatio-temporal correlation construction,the thesis builds a complete spatio-temporal correlation mining algorithm based on spatio-temporal data mining technology,including implicit information transfer,explicit correlation graph construction,graph node differentiation learning,and multi-source graph data fusion,covering the entire process of spatio-temporal correlation mining.The related research has the characteristics of high accuracy,high efficiency,universality,and support for switching between different scenarios in urban spatio-temporal correlation mining applications.The main research content and innovation points are summarized as follows:(1)To solve the problem of unstable city estimation caused by the low data quality of some spatio-temporal nodes,a spatio-temporal meta-learning framework based on probabilistic inference is proposed.It learns the implicit correlation between different spatio-temporal tasks by introducing task-specific spatio-temporal representation and shared statistical structures between different tasks,and transfers spatio-temporal information between different tasks to improve the efficiency of data utilization.(2)To solve the problem of difficult correlation construction caused by inconsistent feature types in urban data,an extension version of mutual information is proposed to measure the correlation between different features,which can handle different types of feature data.The correlation between the obtained data features can be used to construct an explicit correlation structure in the urban system.Based on this,a variational autoencoder based on the correlation structure is proposed to mine the complex correlation in the urban system in an unsupervised manner,and the intermediate products of the model can be used to analyze and trace the urban state to improve the interpretability of the results.(3)To solve the problem of multi-source data fusion and information transmission difficulties caused by heterogeneous data in city system state characterization,a spatio-temporal multigraph meta-learning framework is proposed.It uses graph convolutional networks to learn different information transmission patterns within a single data source based on the node correlations within different data sources,and develops two-partite graph convolutional networks to learn information transmission patterns between non-homogeneous data sources.The fusion of spatio-temporal data features on multiple perspectives in the city enables information transmission,thereby more fully characterizing the city state.(4)To solve the problem of modeling difficulties caused by the complex and changing spatiotemporal patterns in cities,an adaptive dynamic learning model,called spatio-temporal selfbeneficial joint learning model,is proposed.It learns the differences between different spatio-temporal nodes by introducing an additional discriminative task,and balances the information contribution of different nodes in the correlation network through dynamic weighting to enhance the model’s learning ability.The node weights learned in the model can also be used to enhance the interpretability of the model and provide a basis for the data confidence of different nodes.