Identifying multi-regime behaviors of memes in Twitter data

Recent work has studied Twitter's role in distributing information about specific events, in acting as a platform for political debate, and in facilitating social interaction. Despite this interesting body of work, to our knowledge, it is unclear how trending words are used in Twitter, and what is their lifecycle. In this work, we investigate statistical models of the dynamics of word/phrase use in Twitter over time. We identify four base behaviors, derived from the autocorrelation functions of the frequency of word/phrase use. We then observe drift among these base behaviors in our sampled word/phrases over multiple weeks. To the best of our knowledge, this is the first time a hybrid statistical model using Markov processes and ARIMA sub-models have been used to explain the occurrence of certain n-grams within the linguistic space of Twitter topics. The ultimate objective of this work is to develop a hierarchical model for the behavior of word/phrase occurrence within Twitter. The model supposes that words/phrase dynamics move from one regime to another as various exogenous forces act on the population of users. This paper takes the first steps in illustrating that these regimes exist and shows some of the dynamics of regime change.