Early prediction and characterization of high-impact world events using social media

On-line social networks publish information about an enormous volume of real-world events almost instantly, becoming a primary source for breaking news. Many of the events reported in social media can be of high-impact to society, such as important political decisions, natural disasters and terrorist actions, but might go unnoticed in their early stages due to the overload of other information. We ask, is it possible to clearly and quickly identify which of these news events are going to have substantial impact before they actually become a trend in the network? We investigate real-world news discussed on Twitter for approximately 1 year, consisting of 5,234 news events that are composed of 43 million messages. We show that using just the first 5% of the events' lifetime evolution, we are able to predict with high precision the top 8% that have the most impact. We observe that events that have high impact present unique characteristics in terms of how they are adopted by the network and that these qualities are independent of the event's size and scope. As a consequence, high impact news events are naturally filtered by the social network, engaging users early on, much before they are brought to the mainstream audience.