Buzz in Social Media: Detection of Short-lived Viral Phenomena

Abstract

Social media interaction happens in a broad variety of context and magnitude. The vast majority of posts cause little to no discussion, while some start trends and become viral. We study the virality, explicitly of "Buzzes" - posts that evoke intense interaction over a short period of time, as they have been observed frequently, some- times with severe consequences for individuals and companies in the physical world. Early detection of a Buzz may help mitigate or prevent negative consequences of large scale social media outrage against companies or persons, by giving them a chance to react at an early stage. Collecting a labeled set of over 100,000 posts on Facebook pages, we first explore properties that define a Buzz using logistic regres- sion. This method helps us to interpret the results and derive prac- tical recommendations. We subsequently train classifiers and apply machine learning based classification techniques to demonstrate the potential capabilities of automated prediction. We achieve high recall with moderate precision, where feature boosting on broad feature sets yields the most promising results. Our study reveals that Buzzes are well described by a high num- ber of comments from previously passive users, a high number of likes given to comments, and a prolonged discussion period - properties that can be used to distinguish inconsequential posts from potentially volatile ones.