In the context of the current social data revolution, an important question that emerges is “How much information is being leaked by coarse meta-data like time of posts?” Our work on inference from time shows that detailed per-user inferences can be made from the times when a user posts content. Specifically, we see that the correlation between the times when a topic trends on twitter and the times when a particular user tweets can inform us of the user’s interest in the topic. We formulate this interest inference problem as a binary hypothesis-testing problem: users interested in the topic tweet more frequently when the topic trends than at other times & others do not. For example, we have been able to distinguish fans of a baseball team from non-fans using just tweet times of the users and the game times of the baseball team.

We are currently investigating how well these topic-specific trending times can be learned from aggregate twitter activity using coarse information - the time of posts that match a seed keyword associated with the topic.


Dinesh Ramasamy, Sriram Venkateswaran


Upamanyu Madhow