Research And Application Of Regression Prediction Problem Based On Temporal Graph Neural Network

In the real world,many regression prediction problems can be modeled in a graph structure,the data entity variables are represented as nodes on the graph,and the connectivity or correlation among the data entity variables can be used as the potential link relationship among those variables.At the same time,in most cases,the characteristics of nodes in the network dynamically change over time.It not only makes the nodes in the network have temporal change characteristics but also makes the correlation among nodes of different spatial positions in the network have temporal related characteristics in the time dimension.The focus of this thesis is to research on the common problem existed in the regression prediction problem of graph network with temporal characteristics.The main work is as follow:(1)Aiming at the node level regression prediction of the temporal graph of static topology type,this paper proposes a temporal graph neural network model suitable for static topology.Initially,in view of the high-order spatial feature modeling problem of the node itself,a graph convolution module combining Node2 vec and GCN is proposed to capture spatial feature information more comprehensively.Moreover,aiming at the dual influence learning problem of node features and topological spatial structure in a short period of time,a bidirectional temporal spatial encoding layer with residual graph attention is suggested.Finally,in order to improve the influence of the network state at the critical moment on the prediction moment and to alleviate the error propagation of the model,a temporal multi-head attention layer is introduced to solve this problem.(2)Aiming at the regression prediction of nodes on a temporal graph of dynamic topology type,this paper proposes a temporal graph neural network model suitable for dynamic topology.Firstly,the temporal dependence learning method on the temporal graph is improved by using the improved temporal information encoding module.Furthermore,in terms of the problem of multi-type nodes and links modeling in the temporal graph,an improved type-based multi-head graph attention mechanism is proposed to integrate tempral features into the process of graph convolution to realize the automatic capture of important semantic information in the network.(3)Based on a temporal graph neural network suitable for static topology,a traffic speed prediction method is implemented.Experiments conducted in two real-world traffic datasets verify the model's ability to model spatial and temporal dependencies.Finally,an application system prototype is employed to present the effect.(4)Based on temporal graph neural network suitable for dynamic topology,an academic literature citation prediction method isconducted.Besides,an imporved heterogeneous subgraph sampler focused on citation sequence is proposed to extract the neighborhood topology information corresponding to the citation sequence and dynamically assign time labels to nodes,so that the model can preserve the rich semantic information on the graph and be extended to large-scale temporal graph networks.At the end,the effectiveness of the proposed model is verified by experiments and the effect is presented by an application system prototype to present the effect.