Jeudi 20 mars 2014 à 14h, salle 25-26/101
New products, ideas, norms and behaviors are often thought to propagate through a person-to-person diffusion process analogous to the spread of an infectious disease. Until recently, however, it has been prohibitively difficult to directly observe this process, and thus to rigorously quantify or characterize the structure of information cascades. In one of the largest studies to date, we describe the diffusion structure of billions of events across several domains. We find that the vast majority of cascades are small, and are characterized by a handful of simple tree structures that terminate within one degree of an initial adopting « seed. » While large cascades are extremely rare, the scale of our data allows us to investigate even the one-in-a-million events. To study these rare, large cascades, we develop a formal measure of what we label « structural virality » that interpolates between two extremes: content that gains its popularity through a single, large broadcast, and that which grows via a multi-generational cascade where any one individual is directly responsible for only a fraction of the total adoption. We find that the very largest observed events nearly always exhibit high structural virality, providing some of the first direct evidence that many of the most popular products and ideas grow through person-to-person diffusion. However, medium-sized events — having thousands of adopters — exhibit surprising structural diversity, and are seen to grow both through broadcast and viral means. Finally, we show that our empirical results are largely consistent with an SIR model of contagion on a scale-free network, reminiscent of previous work on the long-term persistence of computer viruses.