Spring 2013 Seminar Series

DATE: February 14th
LOCATION: E51-315
TIME: 4:15pm
Reception immediately following

SPEAKER:
Tauhid Zaman

TITLE
Predictive Models of User Behavior in Twitter

ABSTRACT
Are people predictable? Does their behavior follow a certain model? In this talk we will see that within the context of Twitter, the answer is yes. Twitter is a micro-blogging site where users post messages known as tweets which are spread by an action known as retweeting. We propose two models which capture the structural and temporal dynamics of users’ retweeting behavior on Twitter. First, we propose the Superstar Model to describe the structure of “retweet graphs”. This model is motivated by the observation that retweet graphs possess a single “superstar” vertex with degree on the order of the graph size, a phenomenon not modeled by most common random graph models. Our Superstar Model predicts a precise relationship between the superstar degree and the degree distribution of the non-superstar vertices. We observe this relationship in actual Twitter retweet graphs covering a wide array of topics. Our analysis of this new Superstar Model utilizes the theory of multi-type branching processes along with a novel surgery procedure on the retweet graph and is interesting in its own right. Second, we propose a Bayesian model for the continuous time evolution of the retweet graph for an individual tweet. With this model we are able to predict how many retweets a tweet receives using only basic information about the users and the timing of the retweets. Within minutes of a tweet being posted we are able to accurately predict the number of retweets. This model is fairly general and can be extended to user generated content in other social networks. This predictive capability opens up the possibility of placing dynamic display advertisements within such content and compensating users accordingly.