Semi-supervised dictionary learning for network-wide link load prediction

Being a primary indicator of network health, link traffic volumes are used in multiple network management and diagnostic tasks. Although link volumes are available using off-the-shelf tools, the corresponding measurement records typically contain errors and missing data. To overcome these challenges, the present paper develops a link traffic prediction algorithm that fills missing entries and removes noise from the observed entries in an online fashion. The algorithm not only exploits topological knowledge of the network, but also learns from the available historical link traffic data. During its operational phase, the novel algorithm relies on a sparse signal representation for the link counts over a data-driven dictionary. Prediction of link counts follows after solving an ℓ1-regularized least-squares problem. Prior to operation however, a dictionary is trained so that it captures all the necessary information from the historical data, allows for a sparse representation, and is aware of the network topology. This is accomplished through a novel semi-supervised dictionary learning scheme which works even when the training data has missing entries. Numerical tests on data from the Internet2 archive corroborate the proposed algorithms.