Research On Spatio-temporal Graph Data Prediction Model

As cities tends to be intelligent gradually,the rapid development of information technology and mobile sensor networks,a wealth of spatio-temporal data was generated in people’s social activities.Spatio-temporal data is a kind of data with spatial,temporal and attribute features at the same time,and complex structural information exist both spatial and temporal dimension.As a result,spatio-temporal data are mostly represented in the form of graphs.How to exploit the complementary characteristics and discover the value hidden in such data with graph structure to serve the tasks in real-world situations is a tough problem that attracts concerns from both academic and industrial fields.Spatio-temporal data has the characteristics of multi-dimension,multi-granularity,dynamic evolution,etc.,and has complex correlations in time,space,and semantic attributes.Compared with traditional data mining,spatio-temporal graph data mining is more challenging,especially in prediction problems that the characteristics of node imbalance in spatial distribution,variable data evolution,complex node semantic meaning and dynamic graph structure of spatio-temporal graph data pose great challenges to the design of data prediction models.Aiming at these challenges,this dissertation provides multi-correlation partition in spatial dimension,time series modeling,semantic attributes constraint centered on spatiotemporal nodes,and interpretable spatio-temporal graph modeling centered on spatio-temporal graph.The main contributions are as follows:(1)To tackle the problem of spatio-temporal graph data prediction in the case of unbalanced spatial distribution,a new prediction model via multi-correlation modeling is proposed.By exploiting the essential characteristic of spatio-temporal nodes,static graph channel and dynamic graph channel are designed.For static graph channel,an edge weight updating mechanism is designed,which combines the idea of bilateral filtering,considering the spatial features similarity and attribute features similarity between spatio-temporal nodes at the same time.For dynamic graph channel,an edge weight updating mechanism with dynamic auxiliary temporal features is designed,which can satisfy the demand of dynamic change of spatio-temporal graph.The results of experiments in air quality prediction task show that the proposed model outperforms existing prediction models by constructing static and dynamic spatial graphs.(2)To tackle the problem of spatio-temporal graph data prediction in the case of variable temporal characteristics,a new prediction model based on the fusion of long and short term features is proposed.It improves the generalization of short-term feature representation of spatio-temporal nodes by an Auto Encoder based reconstruction model,and uses Memory Network for long-term patterns storage and update.Then an adaptive edge weight updating mechanism is designed to learn dynamic one-step spatio-temporal graph.In addition,the proposed model take the principal temporal feature of fine-grained and coarse-grained spatiotemporal nodes as the main and auxiliary task,respectively.As a result,the multi-task joint optimization function is established to further improve the prediction accuracy of the main task.The results of experiments in air quality prediction task show that the proposed model can effectively learn the short-term features and long-term patterns,and outperforms existing prediction models.(3)To tackle the problem of spatio-temporal graph data prediction in the case of spatial feature evolution with semantic constrained,a new prediction model based on semantic evolution analysis is proposed.This model fully considers the temporal and semantic attribute features,and proposes a novel algorithm with time constraints and the concept of completion ratio of spatial sequence for spatio-temporal node similarity measurement,which improves the accuracy of existing clustering-based prediction methods.Moreover,by exploiting the pretrained semantic transition probability matrix,an adaptive semantic optimization scheme is designed to further filter candidate positions in each cluster.The results of experiments in destination prediction task show that the proposed model can obtain more reasonable candidate destinations,and outperforms existing prediction models.(4)To tackle the problem of poorly interpretability of one-step spatio-temporal graph modeling,a new prediction model based on external knowledge is proposed.It explicitly defines the local features,global features and high-order features of spatio-temporal node simultaneously.Then,an adaptive edge weight updating mechanism integrating domain knowledge is designed to model the spatio-temporal graph in different time steps,which improves the rationality and interpretability.The results of experiments in air quality prediction task show that the proposed model can effectively learn the one-step spatio-temporal graph with Gaussian diffusion model,and outperforms existing prediction models.