Research On Accurate Recovery Of Missing Network Measurement Data With Localized Tensor Completion

With the rapid development of computer network technology,amounts of data are generated all the time.In various engineering applications,these data are needed to in some scientific research and commercial applications.Since the cost of collecting all the data is too great or the data is lost during the transmission process,the obtained data is often incomplete,but these incomplete data still have great research and analysis value.In many network projects such as anomaly detection,network performance analysis,computer vision,and social network analysis,the task of inferring all data through partially monitored network data is becoming more and more important.At present,research in related fields shows that tensor completion is an effective method to make more accurate inference of missing data by using multidimensional data structures.However,the existing tensor completion algorithm generally assumes that the modeled tensor data has the property of global low rank,and tries to find a single overall model to fill the entire tensor data,which ignore the local correlation between data.The assumption result in that the accuracy of the inferred data is too low.In the actual network traffic data,there is a strong correlation between local data,and the accuracy of recovery data can be improved by using this property.In this paper,the Local Tensor Completion algorithm(LTC)is proposed to achieve more accurate data recovery by using the low rank of the sub-tensor.Although this technology is very promising,it still faces two challenges.On the one hand,how to extract the data with strong correlation from the tensor data to construct the local low-rank tensors;on the other hand,how to use the sub-tensor data which is completed to recover the original tensor.In this paper,the related solutions are proposed for these two challenges in the local tensor completion algorithm are as follows:(1)How to extract the local low-rank tensor in the LTC algorithm.In this paper,a kind of selection algorithm for candidate anchors based on local sensitive hash(LSH)is proposed.Using the strong correlation between local data,several low-rank sub-tensors are constructed with the anchor point as the center,and then the local sub-tensor is completion.Experiments show that the candidate anchor selection algorithm based on LSH has faster convergence speed and better effect than the random selection anchor algorithm.(2)Aiming at the problem of finding similar data in sparse tensor data,a new strategy based on slice coding is proposed in LTC algorithm and the definition of similar distance is given.CANDECOMP/PARAFAC decomposition is used to realize sparse tensor,the coding of volume data overcomes the challenge of calculating distance difficulties.(3)For the LTC algorithm,how to complete the original tensor with the sub-tensor data which have been recovered,in this paper,the algorithm of similar sensitive data fusion is given,and the sub-tensor data is weighted to recover the original tensor;At the same time,the anchor point selection algorithm based on sampling density and distance is given to further improve the accuracy of local low-rank tensor completion and reduce the computational complexity of the algorithm.The experimental results show that the proposed LTC algorithm model can effectively improve the accuracy of data recovery compared with other low-rank hypothesis tensor completion algorithms such as CP nmu,CP opt,and CP als.