Title:  "Understanding Cautious Collective Classification"
Speaker: Luke McDowell, U.S. Naval Academy

Traditional machine learning algorithms for classification assume that
instances are independent from each other. In practice, however,
instances are often implicitly or explicitly linked (e.g., by
hyperlinks, email messages, or common interests). Recent algorithms
for "collective classification" (CC) attempt to improve accuracy by
exploiting such links and by making simultaneous inferences about the
instances. To do so, these algorithms rely on intermediate instance
label predictions, which can be incorrect. If ignored, these errors
can reduce or even reverse CC's performance gains.

In this talk, we describe how "Cautious Collective Classification" can
address the problem of uncertain intermediate predictions. We describe
the need for cautiously utilizing such predictions, and explain how
Cautious CC can improve performance without substantially increasing
computational complexity. Then, using a combination of real and
synthetic data, we identify data characteristics for which Cautious CC
yields especially significant accuracy gains.

Bio:

Luke McDowell received his bachelor's degree in Electrical Engineering
from Princeton University in June 1997.  After working on
high-performance image processing at Sarnoff Corporation in Princeton,
NJ, he moved to the Department of Computer Science and Engineering in
Seattle, WA, where he received a Ph.D. in August 2004. His
dissertation focused on making the Semantic Web practical for
"ordinary" people, and was nominated by his department for the 2004
ACM Distinguished Dissertation Award.

In January 2005, McDowell moved to the U.S. Naval Academy in
Annapolis, MD, where he is currently an Assistant Professor in the
Computer Science Department. His research focuses on using machine
learning and text mining techniques to make sense of large amounts of
information, especially the vast quantities of text that can be found
on the web.