| With the proliferation of mobile devices and the diversification of applications,people are becoming more and more connected to wireless networks.This rapid growth in user demand has led to an increase in the size of networks.Thus,posing many challenges to current network infrastructures.Operators are having to deploy more base stations and data processing units to accommodate the growing demand which will significantly increase operating costs.Accurate prediction of future traffic data can optimize network resource scheduling,improve resource utilization,reduce operating costs and ensure network quality.Therefore,it has become a common topic of discussion in academia and industry.In addition to the field of wireless networks,traffic prediction is also of great importance in the field of transportation.Accurate traffic forecasting can ease traffic congestion and reduce accident rates.It can also improve the efficiency of taxi resource allocation to facilitate travel.Traffic data is emergent and non-smooth.Traditional prediction models can no longer accurately characterize them,so this paper uses a graph neural network-based approach to the traffic prediction task.The main elements are as follows.To address the complex spatial correlation of network traffic,this paper proposes to extract the spatial features of the traffic data using a graph convolutional neural network modified based on an attention mechanism.A convolutional network with multi-scale convolutional kernels is also used to extract the temporal features of the traffic.Experimental results on both datasets demonstrate that the network structure proposed in this paper achieves the best performance on a variety of evaluation metrics.To address the problem that graph neural networks are difficult to handle large-scale traffic data,this paper proposes a migration learningbased traffic prediction method.The method first partitions large graphstructured traffic data into several sub-graphs,then uses an unsupervised graph classification algorithm to determine the migration direction,and finally completes the model migration.Experimental results on real traffic data show that the migration learning strategy proposed in this paper can significantly improve the model fitting speed and solve the large-scale traffic prediction problem.To address the characteristic of dynamic spatial correlation of traffic data,this paper proposes a traffic prediction algorithm based on dynamic graph neural network.Based on the spatial correlation at different times,a dynamic spatial structure graph is constructed,which can aggregate spatial information more accurately,overcoming the problem that static graphs tend to introduce noise that makes information aggregation inaccurate.In addition,the graph attention network is used with the power to aggregate spatial information in a weighted manner and a gated recurrent neural network is used to extract temporal information.Experiments on real traffic datasets demonstrate that the algorithm based on dynamic graph neural networks outperforms traditional static networks. |
