Structural Virality

What it is

Every piece of content that spreads can be drawn as a tree. The first person to post it is the root, and each later person who adopted it is attached to the earlier person they most plausibly got it from. A tree in which everyone is attached directly to the root is a pure broadcast. A tree in which people are attached to people who are themselves attached to others, generation after generation, is a chain of person-to-person sharing.

The authors define structural virality as the average distance between all pairs of people in that tree. In a pure broadcast the average distance is about two, because any two adopters are linked through the single source. As the tree branches into more generations the average distance rises. A sprawling, multigenerational tree scores far higher than a broadcast of the same size, so two pieces of content with the same reach can sit at opposite ends of the scale.

The 2016 paper applied the measure to tweets containing links to popular news, video, image and petition sites, collected over the twelve months from July 2011 to June 2012. The authors rebuilt each tree by linking every adopter to the most recent earlier adopter among the accounts they followed. The abstract puts the scale at about a billion diffusion events.

In effect

The main result is that large audiences did not usually come from long chains. Across news stories, videos, images and petitions, popular events spread through a mix of broadcast and viral mechanisms, in every domain and at every size. Structural virality was generally low and stayed low as events grew, so popularity depended mainly on the size of the largest broadcast.

The detail in the full text points the same way. Most cascades were tiny and ended within one generation. Trees with at least 100 members were about 0.025 per cent of the total, roughly one in four thousand. Among those larger trees, the link between size and structural virality was weak: the correlation was about 0.2 for news, about 0.04 for petitions, and essentially zero for pictures and videos. For images and videos the median structural virality stayed under three even for cascades of ten thousand or more reposts.

The measure also changes how other findings can be read. Berger and Milkman found in 2012 that positive and high-arousal articles were more likely to reach a newspaper's most-emailed list, after controlling for attention such as home-page placement. Structural virality offers a way to ask a further question about any such content: once it was seen, did it travel onward through many hands, or did most of its audience come from a single large source? A related 2014 study of photo reshares reported that, early in a cascade, breadth was a better signal of eventual size than depth, which is consistent with broadcast doing much of the work.

The idea sits close to Cumulative Advantage. That concept describes how early visible success draws more success, so that small accidental leads widen. Structural virality describes the shape of the spread that results. One is about why gaps open between hits and the rest, and the other is about whether the hit travelled through a few large broadcasts or many small passes.

What it does not say

It does not say that person-to-person sharing is unimportant, or that content has no part in what spreads. The authors found a mix of mechanisms at every size, and structural virality varied widely among events of similar size. The measure is descriptive: it records the shape of a spread after the fact and does not identify the cause of that shape.

The findings come from one service and from web content linked in posts. The authors state that spread through other channels, such as private messages, falls outside the data and could change the reconstructed trees, and that their attribution of each adopter to a source is imperfect. A check using hashtags gave qualitatively similar results.

The authors also tested a simple model of low-rate contagion on a scale-free network. It matched several of the empirical findings but did not reproduce the diversity of structural virality seen in the data, which leaves the question of what generates the full range of shapes open.

Finally, the measure is not a forecast. It describes events that have already happened and does not tell a creator what to publish. Its practical use is to sharpen the question: before asking what made a piece spread, ask which route its audience took.

---