Information Dynamics in Social Interactions: Hidden Structure Discovery and Empirical Case Studies
Aug 13, 2013
from 10:00 AM to 12:00 PM
|Where||Engr. IV Bldg, Faraday Room 67-124|
|Contact Name||Zicong Zhou|
|Add event to calendar||
Advisor: Professor Vwani P. Roychowdhury
As collective human activity and knowledge continues to be digitized and stored, it provides an unprecedented opportunity to understand what topics are important, how they evolve, and how individuals and organizations interact to form groups and make decisions. The petabytes of data collected everyday, however, underscores the need for new computational tools to help organize and understand these vast amounts of information. The focus of this dissertation has been to develop such tools, and present empirical case studies that both establish the efficacy of the developed tools, and provide new insights into the data sets themselves. For example, (i) We analyze a publicly accessible movie database and find global patterns in the underlying collaboration dynamics, and then show how such emergent patterns can be generated from stochastic decisions made at the level of the actors; (ii) We analyze the so called Twitter revolution that was precipitated by the 2009 elections in Iran, and determine a model for the spread of news on Twitter; and finally, (iii) We develop a novel methodology for Topic Modeling, where given a large corpus of documents, it automatically infers the underlying topics and computes a distribution of documents over the computed topics. Our approach is very different from the highly popular and widely used existing topic models: Instead of using a bag of words model, it is inspired by how knowledge is organized in our brains as an associative network, and it exploits the idea of source coding from information theory to infer the latent networks directly from text data. We apply our algorithms on large-scale corpuses, and using automatic evaluation techniques, show that our topic organization is not only more coherent semantically, compared to the state-of-the-art Latent Dirichlet Allocation (LDA) results, but is also computationally more efficient.
Zicong Zhou is a Ph.D. candidate at the University of California, Los Angeles, supervised by Professor Vwani Roychowhury. He received his B.S. degree in Electronic Engineering from Zhejiang University, Hangzhou, China in 2007, the M.S. degree in Electrical Engineering from UCLA in 2009. In the summers of 2011 and 2012, he was a data scientist intern at LinkedIn and Twitter, respectively, working on large-scale machine learning problems. His research is in the area of large-scale data mining specialized in graph mining and text mining.