Network Traffic Spatial-Temporal Modeling And Forecasting Based On Graph Neural Networks

With the rapid development of urbanization and the economic level,the number of vehicles is rapidly increasing.It leads to existing transportation infrastructures cannot meet the requirement of society.Hence,traffic congestion has become a major problem for many urban and highway networks.Around the world,Intelligent Transportation Systems(ITS)could drastically solve this problem by combining many advanced technologies,such as information and sensor technologies.As an important step of ITS,accurate spatial-temporal modeling and forecasting can effectively improve the level of traffic management,travel routing,and other applications.This paper aims at solving some existing drawbacks of traffic parameters forecasting and congestion forecasting.To this end,this paper proposes four novel graph-based neural networks to achieve more accurate and promising forecasting.The main contributions of this paper are listed below.1.Graph-based deep learning methods work well for traffic parameters forecasting.These methods usually highly rely on predefined graph structures,which are constructed based on the structure information of road networks.However,the quality of the predefined graphs sometimes is not satisfactory.Even the predefined graphs are not available at times.In addition,the predefined graph structures usually only can represent static spatial structures.And they cannot model dynamic spatial relations from road networks.To overcome the above-mentioned drawbacks,this paper proposes an Attention-based Dynamic Graph Generation(ADGG)module.And this paper further designs a novel spatial-temporal graph neural network for traffic parameters forecasting.The proposed algorithm can generate dynamic graphs based on traffic data without the predefined graphs and has a better performance in modeling dynamic spatial relations.Experimental results on two real-world datasets show that,compared with other methods,this work can achieve comparable or better performance without predefined graphs.2.In traffic parameters forecasting,it is hard for previous graph-based methods to fully model complex relations among multiple nodes.Traditional graphs use edges to IV describe correlations among nodes.Each edge only can represent one-to-one relations between two nodes.However,complex many-to-many correlations widely exist in traffic road networks,which are difficult to be represented by edges.To address this problem,this paper introduces directed hypergraphs to represent complex spatial relations from road networks.Directed hypergraphs leverage directed hyperedge to encode many-to-many relationships among nodes.By combining attention mechanisms and directed hypergraphs,this paper further proposes a novel directed hypergraph attention network.Compared with graph-based methods,this work has an obvious advantage in modeling many-to-many relations.Experimental results prove the effectiveness of this work.3.In traffic congestion forecasting,autoencoder and its variants are widely applied for feature extraction.However,the traditional autoencoder cannot model spatialtemporal dependencies in traffic data.And the traditional autoencoder is an unsupervised method.Hence,it may extract useless features for congestion forecasting at times.To address this shortage,this paper proposes a novel autoencoder,Spatialtemporal Graph Discriminant Autoencoder,for traffic congestion forecasting.This work combines recurrent neural networks and graph convolution networks to design an encoder and decoder.In this way,the proposed method can model complex spatialtemporal correlations in traffic data.In addition,this paper introduces a distance penalty into the loss function to extract more helpful features for traffic congestion forecasting.Compared with traditional autoencoders and other well-known methods,many experimental results show that this work has a better performance.4.This paper proposes a multitask-learning model for meeting the requirement of multiple traffic forecasting tasks.The proposed model can simultaneously achieve promising traffic parameters and congestion forecasting.Most of the previous related works focus on one traffic forecasting task.They cannot leverage relations among traffic forecasting tasks to improve their performance.There are many well-known multitask-learning architectures in other fields.Due to the complexity of spatialtemporal dependencies in traffic data,these multitask-learning architectures cannot work well for traffic forecasting tasks.To overcome this shortage,this paper designs a novel multitask-learning model,Spatial-temporal Graph Multi-gate Mixture-of-expert.The proposed method utilizes an exclusive expert network and gate network for each task.And all forecasting tasks share some shared expert networks.Each expert network directly extracts spatial-temporal features from traffic data.And each task’s gate network would decide the weights of different shared expert networks.Based on the weighted sum,there is a tower network for each task to make the final prediction.In addition,this paper integrates graph convolution and attention into submodules to model complex spatial-temporal relations from traffic data.In this model,each task can leverage shared experts to enrich information for achieving more accurate predictions.And each exclusive experts focus on its forecasting task.To evaluate the performance of the proposed model,this paper conducts extensive experiments on two real-world datasets for three traffic forecasting tasks(flow,speed,and congestion).Experimental results show that the proposed model can simultaneously achieve promising predictions for three tasks.