Abstract : Online Platforms dedicated to social networking host new social phenomenons. Thus several keywords may suddenly take an unprecedented importance, reflecting the number of dis- cussions they have raised within a short time period. Such bursts in topic discussions are usually referred to as buzz events. We address in this paper the problem of predicting the activity volume associated to a given keyword without a priori knowledge on the underlying social network. To do so, we propose to define social netowrk on a content-centric way. Our approach is evaluated at "industrial scale" on two different social networks: Twitter, a platform with extremely fast dynamics (Kwak et al., 2010), and Tom's Hardware, a worldwide forum network focusing on new technology. The experiments conducted reveal that it is possible to predict activity volume associated to a keyword in social media with high accuracy.