Reconsidering the Foundations of Network Sampling

Recently, there has been a great deal of research focusing on the development of sampling algorithms for networks with small-world and/or power-law structure. The peerto-peer research community (e.g., [7]) have used sampling to quickly explore and obtain a good representative sample of the network topology, as these networks are hard to explore completely and have significant amounts of churn in their topology. For collecting data from social networks, researchers often use snowball sampling (e.g., [2]) due to the lack of access to the complete graph. Leskovec et al. have developed Forest Fire Sampling, which uses a hybrid combination of snowball sampling and random-walk sampling to produce samples that match the temporal evolution of the underlying social network [5]. Hubler et al. have developed a Metropolis algorithm which samples in a manner designed to match desired properties in the original network [3]. Although there has been a great deal of research focusing on the the development of sampling algorithms, much of this work is based on empirical study and evaluation (i.e., measuring the similarity between sampled and original network properties). There has been some work (e.g., [4, 8, 6]) that has studied the statistical properties of samples of complex networks produced by traditional sampling algorithms such as node sampling, edge sampling and random walks. However, there has been relatively little attention paid to the development of a theoretical foundation for sampling from networks—including a formal framework for sampling, an understanding of various network characteristics and their dependencies, and an analysis of their impact on the accuracy of sampling algorithms. In this paper, we reconsider the foundations of network sampling and attempt to formalize the goals, and process of, sampling, in order to frame future development and analysis of sampling algorithms.