Twitter's Big Hitters Seminar

I will present volume-across-time data from several Twitter conversations. These `who-tweeted-who' events may be viewed as time-dependent entries in a large, sparse, interaction matrix. From a network science perspective, there are many interesting computational challenges in monitoring, summarising, modeling and predicting the local and global patterns of behaviour. My focus for this talk concerns the spikes of Twitter activity that can be attributed to high profile events or news items (for example, a Premier League red card or a reality TV voting result). In such cases, we typically observe that the initial spike in bandwidth decays with a recognisable half-life of around 20 minutes. These dramatic, but short-lived, bursts of interest represent marketing opportunities for suitably agile players (as demonstrated by the cookie company Oreo in the 2013 Superbowl).

To understand the nature of the spike-and-decay activity, we propose a discrete time model that takes account of the specific nature of Twitter communication: a fixed underlying tweeter-follower network forms the backbone of a dynamic message-passing process. Our model predicts that a certain network centrality measure plays a key role in describing the spread of information, and we will test this conclusion on real Twitter data.