When Everything Looks Like a Nail: Graph Models of the Internet
Seminar Room 1, Newton Institute
AbstractThe general appeal of abstracting real-world networks to simple graphs is understandable and has been partly responsible for fueling the new field of “network science”. However, as the Internet application has demonstrated, such abstractions that ignore much of what engineers consider as critical come at a price. For example, they can lead to the study of graph models that have little in common with the real-world networks that motivated these models in the first place. In turn, they tend to focus on cyber-threat models that are largely of academic interest because they are incapable of dealing with the most potent and potentially lethal real-world threats.
Fortunately, the Internet application also suggests an alternative and more engineering-based approach to the “art” of abstracting real-world networks. This approach emphasizes network function over network structure and requires (1) a basic understanding of the process by which network connectivity measurements are obtained, (2) a careful assessment of the kind of inferences that the available measurements can and cannot support, and (3) a break with tradition when it comes to network modeling and model validation. I will illustrate with specific Internet-related examples that ignoring any of these issues is bound to lead to specious network models or alluring results that quickly collapse when scrutinized with real data or examined by domain experts.