The Structural Virality of Online Diffusion

• Activity Number: 73
• Type: Topic Contributed
• Date/Time: Sunday, August 4, 2013 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Ashton Anderson and Sharad Goel and Jake Hofman*+ and Duncan Watts
• Companies: Stanford University and Microsoft Research and Microsoft Research and Microsoft Research

Abstract

While theoretical work on modeling the diffusion of products and ideas extends back several decades, it has traditionally been difficult to directly observe such processes at the level of individual adoption events. As such, most past work focuses on more readily available population-level data, leaving open the question of how items spread from one individual to the next. Here we leverage recently available data comprised of billions of diffusion events on Twitter to investigate the empirical structure of successful information cascades, proposing a measure of "structural virality" that interpolates between two extremes: wide but shallow broadcast events from a small number of large sources and multi-generational cascades where any given source accounts for a small fraction of all adoptions. We find surprising diversity in the structure of cascades, with relatively low correlation between size and structure for medium-sized events. We compare these results to simulations from an SIR model of contagion on a scale-free network, finding largely consistent results over a restricted parameter range, with some important differences.