The Trilemma of Network Analysis

The recent interest in network analysis is caused by the unprecedented accessibility to large datasets: there are huge, publicly available databases on protein-protein-interactions, air transportation, and street maps which easily lend themselves to a network representation. Once a network is created, all types of path-based network analytic measures can be easily applied: typical examples are centrality measures, but also some clustering algorithms and robustness analysis rely on path-based measures. Borgatti has claimed that centrality measures basically simulate dissemination processes of goods which use a certain subset of paths on the given network [1]; they can thus only be used to describe processes which rely on the same type of good and the same subset of paths. Later, Butts pointed out that the results of a chosen network analytic method strongly vary with modeling decisions taken when turning raw data into networks [2]. In this article we combine these two insights to the trilemma of network analysis which states that the network process of interest, the network representation, and the network analytic method cannot be chosen independently. We discuss on two real-world examples in the realm of air transportation networks how to choose a distance based measure with respect to the context of the data, re-computing similar analyses by Guimer'a et al. [3] and Dall'Asta et al. [4]. In both cases, the path-based measures matching the network process of interest change the interpretation of the previous findings, which shows the potential in regarding the trilemma of network analysis.