The work of collective behavior is to understand and how actors behave in a social networking environment. A large population is involved in social media like Facebook, Twitter, Flicker, and YouTube that provide opportunities and challenges to study collective behavior on a large scale. In this paper, we aim to learn to predict collective behavior in social media. Consider for given information about some actors, how can we infer the behavior of unobserved actors in the same network? However, Connections in social media is not homogeneous. A social-dimension based approach which represents the relations associated capture prominent interaction among different actors to show effective in addressing the heterogeneity of connections included in social media. The social media has are normally of large size, involving hundreds of thousands of actors. The scale of these networks entails scalable learning of models for collective behavior prediction. We propose an edge-centric clustering scheme to extract sparse social dimensions. With sparse social dimensions, the proposed approach can efficiently handle millions of actors while describing a comparable prediction performance to other non-scalable methods.