Abstract:
|
We investigate a time series of graphs corresponding to network connections over a period of several months. Each graph corresponds to the connections between pairs of computers that are initiated within the time period (one hour in our experiments). We discuss variants of graph spectral embedding methods and discuss some of the current theory that has been developed, particularly by the team at Johns Hopkins. Spectral embedding applied to a time series of graphs requires a further processing step to align the embedded data for different graphs. Traditionally, this is performed by a Procrustes transformation, which requires observations to be paired between the two time steps. We describe a variant that uses kernel estimators, that requires no pairing of the data, and discuss further variants that combine both methods. These ideas are illustrated on the network data, where a subset of the node(computers) are not resolved to unique IP addresses, and therefore cannot be reliably paired between different graphs. The kernel technique allows all the data, paired and unpaired nodes, to be used in the aligning transformation.
|