Sampling as a Method of Comparing Real and Generated Networks

Abstract

In this paper, we combine the use of sampling methods and a network generator to assess the degree of similarity between real and generated networks. Generative network models provide a tool for studying essential network features. These include, for example, the average and distribution of node degree, cluster coefficient and community size. The aim of the generators based on these models is to create networks with properties close to real networks. Even with a high similarity of global properties of real and generated networks, the local structures of these networks often differ considerably. On the other hand, when the network is reduced by a sampling method, global features of networks are strongly influenced by local structures. In the paper, we compare properties of a real-world network and a generated network and also properties of their small samples. In experiments, we show how the distribution of the properties of individual networks change by using different sampling methods and how these distributions differ for both networks and their small samples.