Research On Spatio-temporal Graph Node Attribute Prediction Based On Weighted Graph Diffusion Convolutional Network

With the widespread application of Io T and big data,various spatio-temporal graph data networks are continuously generating large amounts of spatio-temporal data,such as traffic road networks,air testing,and mine seismic monitoring networks.It is of great significance to effectively explore spatiotemporal data feature patterns,capture the common spatiotemporal dynamic dependencies of spatiotemporal graph data,and realize spatiotemporal graph data prediction for the application in different scenarios.For example,in the field of transportation,traffic flow can be predicted in the future.In the field of air testing,future air quality can be predicted,and in the field of mining earthquake,the vibration trend can be predicted.Therefore,spatio-temporal data prediction has great application significance for human life.Spatio-temporal graph data has spatio-temporal correlation,heterogeneity,and periodicity characteristics.To effectively model spatio-temporal graph data for generic downstream applications and accurately predict spatio-temporal sequences,it is crucial to capture the spatio-temporal dynamic correlations.Modeling approaches for spatio-temporal graph data include dynamic modeling and data-driven methods.Dynamic modeling uses a mathematical modeling approach,which has strict conditional assumptions and requires feature extraction or manual intervention,thus severely limiting the model’s flexibility.Data-driven approaches mostly use statistical learning and machine learning,in which statistical learning mostly considers static time-series data and ignores spatio-temporal correlation,which cannot achieve good learning results.In deep learning methods,researchers combine graph neural networks with temporal learning models to improve the prediction results.However,most of the current research treats spatio-temporal graphs as static graph structures,while actual spatio-temporal evolution is dynamically correlated and subject to a complex diversity of variation due to multiple feature factors.Therefore,to address the above research deficiencies and the difficulty of capturing the dynamic dependence of spatio-temporal graphs,an dynamic weighting diffusion convolutional learning unit is proposed in this paper.The dynamic correlation of spatio-temporal graph data is studied in-depth,and a spatio-temporal graph node attribute prediction model based on the weighted graph diffusion convolutional network is proposed.The research contents and innovations of the paper are as follows:(1)A weighted parameter learning method is proposed to address the problem that existing methods represent the spatio-temporal graph as a fixed graph structure and cannot effectively learn the spatial dynamic dependence.The spatio-temporal graph is modeled as a directed weighted graph,and the Dijkstra and dynamic time regularization algorithms are used to predefine the graph adjacency matrix in terms of explicit graph structure and hidden semantics.The weighted learning parameter matrix and graph attention coefficient matrix are proposed to effectively learn the spatial dynamic correlation.(2)To address the problem that spatio-temporal graph data have complex spatio-temporal dynamic correlations and cannot effectively predict attribute features under single factor scenarios,the Deep Spatio-Temporal Graph Weighted Diffusion Convolutional Network(DST-WDCN)is proposed to achieve spatio-temporal graph attribute prediction under single feature factors.The Gated Diffusion Convolution is proposed,and combining gated diffusion convolution with Gated Extended Causal Convolution forms Spatio-Temporal Convolutional Blocks.Finally,the stacked ST-Block is connected with the hopping layer to construct the DST-WDCN model.(3)To address the challenge that spatio-temporal data often exhibits periodic and complex changes due to multiple factors under multi-factor scenarios,this paper proposes a dynamic weighting spatio-temporal graph convolutional network(MST-WDCN)that incorporates multi-factor features to enable spatio-temporal graph attribute prediction.The proposed approach involves several components.First,the input data is divided by temporal granularity.Second,introducing a periodic learning component that includes a neighborhood-dependent component and a period-dependent component.Third,this paper proposes an external feature learning component that consists of a graph convolution with heterogeneous feature learning layer.Finally,an output component that employs a weighted learning approach is proposed to construct the MST-WDCN model.(4)This paper validates the proposed model using several large real datasets and evaluate its performance using multiple metrics in comparison with a benchmark model.The experimental results demonstrate that the proposed model outperforms the benchmark model in terms of prediction accuracy.This paper also conducts ablation experiments,case studies,and generalization experiments to verify the effectiveness of each component of the proposed model.The experimental synthesis shows that the proposed model exhibits strong robustness,adaptability,and generalization ability and has a wide range of application prospects.