题目：Clustering and Co-clustering with Bregman Divergences 报告人：Joydeep Ghosh
Schlumberger Distinguished Centennial Chair Professor, IEEE Fellow
Dept. of Electrical & Computer Engineering
The University of Texas at Austin, USA 时间：8月8日14:00-15:00
地点：蒙民伟楼404会议室 摘要：
Bregman divergences form a large class of "distance" or loss functions with
certain common properties.
In the first part of this talk, I will show how the simple k-means algorithm,
which is related to a squared Euclidean loss function, can be generalized to
loss functions for all Bregman divergences. Further, we show an explicit
bijection between Bregman divergences and exponential families, and use it to
derive a simple soft clustering algorithm for all Bregman divergences.
Together these two results enable hard/soft clustering of a very wide range of
data types with their corresponding noise models.
In the second part, we introduce the powerful idea of co-clustering and
propose a general framework for performing it under a variety of loss
functions and domain constraints. The minimum Bregman information solution, a
direct generalization of maximum entropy and least squares principles, plays a
critical role in the analysis that leads an elegant meta-algorithm guaranteed
to achieve local optimality. Some applications of co-clustering will be
illustrated as well. 题目：Simultaneous (Co)-Clustering and Modeling for Large Scale Data Mining
Applications 报告人：Joydeep Ghosh
Schlumberger Distinguished Centennial Chair Professor, IEEE Fellow
Dept. of Electrical & Computer Engineering
The University of Texas at Austin, USA 时间：8月10日15：00-16：00
地点：蒙民伟楼404会议室 摘要：
For several challenging data mining applications, it is often possible to
obtain better and more interpretable results by learning separate models (e.g.
classification/regression) on different data segments/clusters, instead of
trying to fit a single complex model to the entire data. In this talk, I'll
show how the partitioning of the data can emerge naturally while the models
are being developed, using a broad conceptual framework. The emphasis will
be on large-scale, dyadic datasets, e.g. customer-products, user-movies,
webpage-ads, etc. I will also show that the overall approach also provides
nice ways of performing active learning, model selection, as well as
identification of the most reliable predictions among the test cases of
interest Professor Joydeep Ghosh's homepage: http://pegasus.ece.utexas.edu/~ghosh/ 简历：
Joydeep Ghosh joined the UT-Austin faculty in 1988 after being educated at IIT
Kanpur, (B. Tech '83) and The University of Southern California (MS, Ph.D).
He is currently the Cullen Professor in Engineering, and a Fellow of the IEEE.
He is the founder-director of IDEAL (Intelligent Data Exploration and
Analysis Lab). His research interests lie primarily in the theory of adaptive
multi-learner systems, intelligent data analysis, data mining and web mining,
and their applications to a wide variety of complex engineering and AI
problems.
Dr. Ghosh has published more than 250 refereed papers and co-edited 18 books.
He has received 12 best paper awards over the years, including the 2005 Best
Research Paper at UT-Austin from the Co-op Society and the 1992 Darlington
Award given for the Best Paper across all publications of the IEEE Circuits
and Systems Society. He was the Conference Co-Chair of Artificial Neural
Networks in Engineering (ANNIE)'93 to '96 and '99 to '03 and Program Co-Chair
for The SIAM Int'l Conf. on Data Mining (SDM'06). He also serves on the
program committee of several top conferences on data mining, neural networks,
pattern recognition, and web analytics every year. Dr. Ghosh has been a
plenary/keynote speaker on several occasions such as ANNIE'06, MCS 2002, and
WORLDCOMP'07, and has widely lectured on intelligent analysis of large-scale
data. He has co-organized workshops on high dimensional clustering (ICDM 2003;
SDM 2005), Web Analytics (with SIAM Int'l Conf. on Data Mining, SDM2002), Web
Mining (with SDM 2001), and on Parallel and Distributed Knowledge Discovery (
with KDD-2000).
Dr. Ghosh has served as a consultant or advisor to a variety of companies,
from successful startups such as Neonyoyo and Knowledge Discovery One, to
large corporations such as IBM, Motorola and Vinson & Elkins. His research
group has been supported by the NSF, Google, ONR, ARO, AFOSR, Intel, IBM,
Motorola, TRW, Schlumberger and Dell, among others. At UT, Dr. Ghosh teaches
graduate courses on data mining, artificial neural networks, and web analytics
. He was voted the Best Professor by the Software Engineering Executive
Education Class of 2004.
|
