Research On The Graph Convolutional Models On Spatial-Temporal Networks And Their Applications

In recent years,graph convolutional networks have attracted lots of attention in academic communities,and have been applied to many fields,such as recommender systems and knowledge graphs.There are lots of spatial-temporal networks data in spatial-temporal data mining.The spatial-temporal networks data refers to the graph structure data that adds time information and spatial information to the traditional topology graph.However,there is still a lack of effective graph convolutional methods tailored to spatial-temporal network data.Spatial-temporal network data are highly complex,exhibiting strong spatial dependencies,temporal correlations,spatial-temporal correlations,and spatial-temporal heterogeneities.Most of the existing methods cannot take these four properties fully into account.Moreover,the traditional graph convolution methods are difficult to deal with the spatial-temporal properties and influence between nodes when applied to the spatialtemporal networks.This thesis focuses on these two problems,proposed Spatial-temporal Synchronous Graph Convolutional Networks and Adaptive Spatial-temporal Synchronous Graph Convolutional Networks to forecast the spatial-temporal networks.The main contributions are as follows.Most of the existing methods only focus on modeling the spatial dependencies and the temporal correlations,they do not directly consider the spatial-temporal correlations that cross both spatial and temporal dimensions.To address that problem,a localized spatial-temporal graph is designed in this thesis,which connects multiple spatial graphs.Then deploying a graph convolutional method on it to capture the spatial dependencies,temporal correlations,and the spatial-temporal correlations simultaneously.To consider the heterogeneities in different areas and different time periods the spatial-temporal network exhibits,the Spatial-temporal Synchronous Graph Convolutional Networks(STSGCN)has been proposed,which uses a multi-component mechanism and can effectively capture the localized spatial-temporal correlations and heterogeneities.When the traditional graph convolutional methods are applied to the spatial-temporal graphs,these methods cannot distinguish the nodes in the spatial-temporal network correctly and the aggregation of neighbors' features is highly dependent on the adjacency matrix.This thesis proposed spatial-temporal embedding and an adaptive matrix two methods to solve the two deficiencies mentioned above,which consist of Adaptive Spatial-temporal Synchronous Graph Convolutional Networks(ASTSGCN).The ASTSGCN model can adjust the spatial-temporal properties of the nodes and the weight of edges in the spatial-temporal graphs adaptively so that it has better generalization performance.Extensive experiments have been conducted on 4 subsets of the open datasets Pe MS from California Freeway.The experiment results prove the effectiveness of the two methods,which can forecast the spatial-temporal network effectively.