Work‎ > ‎Projects‎ > ‎

Circus - detecting events in social circles

posted Sep 27, 2013, 2:12 PM by Botao Hu
Research and Development
Jul 2013 - Sep 2013
Supervised by Jiong Wang
Twitter Inc.

This project is aimed to detect circle events like "#sigir2013" from the tweet stream which 
  • is small that usually cannot be detected by trends
  • recently bursts in 1-2 day time window
  • involved users’ friendship network forms a social circle
  • strongly interacts with each others
I build a “replay” version for detecting event circles, which successfully find event circles from historical tweets. 

Compared to traditional community detection algorithms that are based on the connection structure only, this project is focusing on the interest-driven community. Only the community who are all involved in a bursty event will be extracted. There, through this project, we can obtain an interest graph -- correspondence between an interest/event and the engaged community. 

In contrast to the traditional interest mining methods, the correspondence between an interest and a group of engaged people usually tends to have higher signal quality than the correspondence between an interest and an individual. That is because the individual behavior might be recorded sparsely or noisely or even biasedly, but consistent group behavior tends not to be biased individually and have tendency to show strong group interest and inherent interest connection inside the group. 

The another contribution is that I investigate different information spreading patterns:
  • bursty events that have dense interaction graph -- circle event, e.g. #sigir2013
  • non-bursty events that have dense interaction graph -- group chat, or link farm, e.g. #sixsigma
  • bursty events that have loose interaction graph -- breaking news, e.g. #syriacrisis
  • non-bursty events that have loose interaction graph -- meme, or tv show, e.g. #morningtweet and #numb3rs