Explaining Bursts of Attention on Social Media
Social media are like a giant megaphone for public opinion: they can sway elections, crush a business, or incite mass action on hot-button issues like vaccination and climate change. Researchers studying how a topic grabs “collective attention” have noticed a common feature in social media data: occasional short and seemingly random bursts of high-volume activity. These poorly understood “spikes” are an intrinsic aspect of attention dynamics, says Manlio De Domenico, a network theorist at the Bruno Kessler Foundation (FBK) in Trento, Italy. “They can ignite cascades of messages that dramatically amplify the impact of the [idea being] spread,” he says.
Intrigued by this amplification effect, De Domenico and his colleague Eduardo Altmann of the University of Sydney, Australia, have now crafted a simple network model whose inputs can be varied to predict the frequency and amplitude of attention spikes on Twitter. Speaking at the International Conference on Complex Networks in Tarragona, Spain, De Domenico reported that their model reproduced features of actual Twitter bursts related to major events in culture, science, sports, and religion. He says that understanding the origin of these bursts might help researchers develop tools that block illegitimate forms of manipulating public opinion.
De Domenico and Altmann began developing their model in 2016 when both were at the Max Planck Institute for the Physics of Complex Systems in Germany. The two analyzed millions of Twitter messages with keywords connected to several events, including the election of Pope Francis in 2013, the 2015 Finals of the National Basketball Association, the 50th anniversary celebration in 2013 of Martin Luther King’s “I have a Dream” speech, and the 2016 announcement about the discovery of gravitational waves. In a window of time surrounding each event, Twitter activity surged and then fell off. But the data also revealed a seemingly random series of activity spikes, minutes-long blasts during which tweets, retweets, and replies increased up to a thousandfold.
To model these datasets, the team described the Twittersphere as a network of users with a given distribution of “connectedness”—a.k.a. the number of followers a user has. After a user posts a message, the model allows her followers to react—retweeting or replying to the message with a certain probability. With this basic model, the researchers considered different scenarios for user behavior. For example, they could vary the “bias” of each user—the extent to which they preferentially pay attention to influential users. They could also study the effect of network structure by comparing a homogeneous network (most users have the same number of connections) with a heterogeneous one (users are concentrated into clusters where the number of connections is higher than average).
For each event, De Domenico and Altmann fed a random distribution of initial tweets into their model and then simulated the number of replies and retweets. Comparing the simulated outcome with the real-world data, the duo found that only two factors in their model are needed to reproduce the salient features of attention bursts: the user bias and the type of network structure. Specifically, bursty behavior seems to be a signature of a highly heterogeneous structure like that of Twitter, which has influential hubs and a small degree of separation between any two users.
It may seem obvious that Twitter’s “influencers” and network structure play a role in bursty tweeting. But De Domenico says it’s remarkable that they are the only factors that seem to matter. For instance, he was surprised to find that the burst amplitudes do not depend on “correlations” within the Twitter network. Such correlations occur because influential users tend to be connected to each other (Obama is likely to follow Clinton, for example), an effect that leads to clusters of high-profile people or institutions.
Understanding the complexities of social media attention could lead to ways of tackling a rising problem: social media manipulation. An increasing number of bot accounts—feeds using artificial intelligence to impersonate real people—try to influence public opinion by amplifying certain views online. “Using bots is like paying people to attend a rally to make it look more important than it is,” says De Domenico. He is currently investigating whether models such as his could distinguish spikes generated by bots from genuine ones.
That possibility would dovetail with another of De Domenico’s projects. He recently unmasked the action of bots when Spain held a referendum on the independence of Catalunya. Using machine-learning tools to recognize bots and to classify a message’s meaning, he and his colleagues showed that bots posted as much as 20% of Twitter messages in the lead-up to the referendum. These posts selectively targeted influential people within the pro- and anti-independence groups. And they were crafted to exacerbate social conflict, for instance, by bombarding independentists with stories of police violence against demonstrators.
De Domenico next plans to combine his bot-analysis tools with collective-attention modeling to study the effect of bots on voting. If researchers can disentangle the role of bots in generating social media attention, policy makers may start thinking about countermeasures. “We need to understand and control the use of bots if we want the voice of real citizens to be central in our democratic debates,” says De Domenico.
Matteo Rini is the Deputy Editor of Physics.