Deep Learning-based Network Traffic Prediction

In recent years,with the popularity of smart terminals and application software,the Internet is having an increasingly important impact on all aspects of people’s lives,and the scale of the network and the number of users are growing rapidly.At the same time,this also brings a lot of challenges to the construction of network infrastructure,the allocation of network resources and the security of the network.Network traffic is an important indicator that can reflect the operation status and health of the network,so if the future network traffic can be effectively predicted,the above challenges can be well met.However,existing research on network traffic prediction has some shortcomings,mainly including not fully learning the complex multi-range and multi-level spatial-temporal characteristics among network traffic data,and not considering the situation that network traffic data has a large number of missing values.In order to address these shortcomings,in this paper two deep learning-based network traffic prediction methods are proposed,and the specific research works are as following:A Multi-range Multi-level Spatial-Temporal Learning(M~2STL)model based on graph data is proposed,to learn complex spatial-temporal features in dynamic network environments.In M~2STL,three Spatial-Temporal Aggregation(STA)modules are designed to simultaneously learn three different ranges of spatial-temporal features of the recent,the daily periodic and the weekly periodic data.There are three key technologies proposed in STA:firstly,to learn long correlations in temporal dimension with small number of parameters,a Dilated Conv Net based on 1D-dilated convolutions is proposed.And a Mix-hop GCN Net is also proposed based on graph convolutions to extract spatial features.Furthermore,instead of providing the Mix-hop GCN Net with a pre-fixed adjacency matrix,a Graph Learning Net is proposed in STA to learn multiple adjacency matrices from the shallower level to the deeper level in a data-driven way.Extensive experiments are conducted in this paper on two real world datasets,and validated the prediction performance of M~2STL.A network traffic prediction model GCP-CNN with missing data is proposed,which is based on a tension-complementary algorithm.The network traffic data for prediction and the historical traffic data are first modeled into a same sparse tensor.Then,GCP-CNN uses a generalized CP mapping and a neural network mapping to fully exploit the complex nonlinear spatial and temporal features inside the sparse tensor.Thus,a complete tensor containing the prediction results can be reconstructed from the existing historical observations.In this paper,the prediction performance of GCP-CNN is verified in two cases with 70%and 90%missing rate on two real world datasets,respectively.