Recovery And Measurements Of Internet Traffic Data Based On Tensor Completion

Tensor completion fills in the unknown or missing data when only a small part of the data is observed to get all the tensor data.A large amount of data is organized into a tensor during a tensor filling process.Tensor completion becomes critical for various network engineering network,such as capacity planning,load balancing,path setup,network provisioning,anomaly detection,failure recovery.But due to the network measurement resources,hardware conditions of sensor node,monitoring and communication costs,etc.The influence of objective condition,so at the time of data collection network will face great difficulties,therefore,in the study of reality,will only take a small amount of data,however,in many of the network engineering need is all the data in the entire network.Existing studies have shown that the data with a tensor filling recovery is more accurate than the matrix filling recovery data.In previous studies,however,never considered the change of sampling rate,so you can form a complete tensor,sampling in reality,however,that the sampling rate is will change with the change in the flow in the network.At the same time,at the time of random sampling,due to various constraints,random sampling can only receive part of the data,but in the case of low sampling rate,tensor fill the recovery performance is not very good.Therefore,this paper focuses on two major issues,and the main work and innovation points of this paper are as follows:First,we propose a Reshape-Align scheme to form the regular tensor with data from variable rate measurements.In the actual network,due to the change of sampling rate,we cannot get a standard tensor.In the scenario,we put forward a kind of matrix segmentation algorithm based on time alignment,the rules of use of our time effectively to traffic matrix is decomposed into corresponding to at the same time the same sampling rate of the matrix.Then we use these sub-matrices to form the tensor of the rules.The experimental results show that the reconstructed-alignment method can achieve a good recovery rate when the sampling rate changes.Second,Rather than taking random measurement samples,to further reduce the measurement cost,we propose a pre-scheduled sampling algorithm to actively determine where to take future network monitoring samples while at the same time ensuring the accurate inference of unmeasured data through the tensor completion.It is challenging to find the optimal sampling points without knowing the structure of the future data.To conquer the challenges,our scheduling scheme includes several novel techniques: a three-part graph to represent the tensor data,a graph-based sample selection algorithm with well-designed selection principle and feasibility check procedure,and a graph-based tensor completion algorithm.The experimental results show that the pre-scheduled method can achieve good recovery performance even at very low sampling rate.