Classifying Internet Traffic Using Linear Regression

Abstract

A globally weighted regression technique is used to classify 32 monitoring sites pinging data packets to 513 unique remote hosts. A statistic is developed relative to the line of best fit for a 360° manifold, measuring either global or local phase correlation for any given monitoring site in this network. The global slope of the regression line for the variables, phase and longitude, is standardised to unity to account for the Earth's rotation. Monitoring sites with a high global phase correlation are well connected, with the observed congestion occurring at the remote host. Conversely, sites with a high local phase correlation are poorly connected and are dominated by local congestion. These 32 monitoring sites can be classified either globally or regionally by a phase statistic ranging from zero to unity. This can provide a proxy for measuring the monitoring site's network capacity in dealing with periods of peak demand. The research suggests that the scale of spatial interaction is one factor to consider in determining whether to use globally or locally weighted regression, since beyond one thousand kilometres, random noise makes locally weighted regression problematic.