Challenges to inferring causality from viral information dispersion in dynamic social networks

摘要

Understanding the mechanism behind large-scale information dispersion through complex networks has important implications for a variety of industries ranging from cyber-security to public health. With the unprecedented availability of public data from online social networks (OSNs) and the low cost nature of most OSN outreach, randomized controlled experiments, the "gold standard" of causal inference methodologies, have been used with increasing regularity to study viral information dispersion. And while these studies have dramatically furthered our understanding of how information disseminates through social networks by isolating causal mechanisms, there are still major methodological concerns that need to be addressed in future research. This paper delineates why modern OSNs are markedly different from traditional sociological social networks and why these differences present unique challenges to experimentalists and data scientists. The dynamic nature of OSNs is particularly troublesome for researchers implementing experimental designs, so this paper identifies major sources of bias arising from network mutability and suggests strategies to circumvent and adjust for these biases. This paper also discusses the practical considerations of data quality and collection, which may adversely impact the efficiency of the estimator. The major experimental methodologies used in the current literature on virality are assessed at length, and their strengths and limits identified. Other, as-yet-unsolved threats to the efficiency and unbiasedness of causal estimators-such as missing data-are also discussed. This paper integrates methodologies and learnings from a variety of fields under an experimental and data science framework in order to systematically consolidate and identify current methodological limitations of randomized controlled experiments conducted in OSNs.