Implementing Stochastic Gradient Descent Based On Historical Network Distance For Available Bandwidth

Predicting network bandwidth of large systems based on a few pairs of network nodes is essential to overcome large measurement overhead over full- mesh active measurements. Recently, prediction using low-rank matrix factorization has gained attention. The algorithm is fully decentralized where no explicit matrix constructions or special nodes such as landmarks and central server is needed. Prediction error and convergence to global minimum are two major concerns of this type of algorithm. In this paper, we propose to enhance low-rank matrix factorization by Stochastic Gradient Descent (SGD) initialized with Singular Value Decomposition (SVD). Experimental results show enhanced prediction error and convergence performance is achieved through our approach.