Statistical mechanics of Twitter

From fads to residential segregation, social processes depend on interactions among many individuals. Social phenomena are prototypically emergent, leading some to ask if we can build a statistical mechanics for these systems. We construct such models directly from data on individual participation in Twitter communities. We identify communities and topics of conversation within those communities, allowing the definition of binary (tweet/silent) variables for each individual during each conversation. We then build maximum entropy models that match the pairwise correlations among these variables, predicting the joint distribution of the binary variables across the entire community. These simple Ising-like models give an accurate quantitative description of many higher order features in the data, and lie near a critical surface in the space of possible models. Finally, we systematically coarse-grain the observed network states, finding hints that the macroscopic behavior is controlled by a non-trivial fixed point in the sense of the renormalization group.