Event Location: NSH 1507
Bio: Lexing Xie is Senior Lecturer in the Research School of Computer Science at the Australian National University. She was with IBM T.J. Watson Research Center in New York from 2005 to 2010, and adjunct assistant professor at Columbia University 2007-2009. She received B.S. from Tsinghua University, Beijing, China, and M.S. and Ph.D. degrees from Columbia University, all in Electrical Engineering. Her research interests are in multimedia, social media, and applied machine learning. Her recent projects are on multimedia analysis, social media tracking, visual semantics, large-scale image and video search, geo-spatial and urban event modeling. Lexing’s research has received five best student paper and best paper awards between 2002 and 2011, and a Grand Challenge Multimodal Prize at ACM Multimedia 2012. She was the 2005 IBM Research Josef Raviv Memorial Postdoc fellow in Computer Science and Engineering. She is an associate editor of ACM Transactions on Multimedia Computing, Communications and Applications, she regularly serves on the program and organizing committees of major multimedia conferences.

Abstract: Multimedia is growing to take up more than 50% of the internet traffic. Understanding these content and their social traces presents new research challenges and opportunities at the intersection of image and video understanding, text analysis, and mining the social web. I will first give an overview about recent work in my group on analyzing real-world event traces in social media, including: use hyperlink patterns to diffusion flow about news events, track visual memes on Youtube, and analyzing rich-media microblogs with cross-media topic model. The second half of the talk will address the question: “which thousand words describe a picture?” We make a novel connection among three distinct large online resources of social tagging, labeled pictures, and commonsense knowledge. We extract statistics about the visual informativeness of picture tags and tag-tag relationships from 5 million photos and 20,000 tags. We propose a model to infer latent network relationships, which is optimized as an inverse graph ranking problem. We use these observation to improve image tagging, and obtained good results on a dataset in the order of 100,000 pictures.