Low-Rank Representation for Internet Traffic Reconstruction Using Compressive Sampling

Abstract

We study compressive sampling for internet traffic reconstruction. Compressive Sampling (CS) requires that the traffic satisfies the low-rank feature. Low-rank states that traffic matrix can be represented in the right domain which the entire necessary information is concentrated in a low number of coefficients. In this paper, we compared three low-rank representation, which are Principal Component Analysis (PCA), Singular Value Decomposition (SVD), and Singular Value Decomposition Mean (SVDM). This low-rank representation is applied to four CS reconstruction algorithms, namely: Sparsity Regularized Singular Value Decomposition (SRSVD), Singular Value Decomposition L1 (SVDL1), Iteratively Reweighted Least Square (IRLS), Orthogonal Matching Pursuit (OMP), and Interpolation. The SVD outperforms the others low-rank representation techniques when used together with SRSVD, SVDL1, IRLS, and Interpolation. The SVDM gives the best NMAE when applied to the OMP. The computational times is linear with the number of the rank matrix. For all reconstruction algorithms, SVDM takes the least computational times.