There is a signup schedule for this event.

Speaker Na

me/Affiliation: Paul Bennett/Carnegie Mellon University

Date/Time:

Friday August 25 2006 at 11:00 a.m.

Location: ACES 6.304

Ho

st: Ray Mooney

Talk Title: Building Reliable Metaclassifiers for Tex

t Learning

Talk Abstract:
Appropriately combining information
sources is a broad topic
that has been researched in many forms. It inc

ludes sensor
fusion distributed data-mining regression combination
classifier combination and even the basic classification
problem. Afte

r all the hypothesis a classifier emits is just a
specification of how

the information in the basic features
should be combined. This talk add

resses one subfield of this
domain: leveraging locality when combining c

lassifiers for text
classification. Classifier combination is useful in
part as
an engineering aid that enables machine learning scientists to

understand differences in base classifiers in terms of their
local r

eliability dependence and variance -- much as
higher-level languages a

re an abstraction that improves upon
assembly language without extending
its computational power.

After discussing and introducing improved

methods for
recalibrating classifiers we define local reliability
d

ependence and variance and discuss the roles they play in
classifier co

mbination. Using these insights we motivate a
series of reliability-in

dicator variables which intuitively
abstract the input domain to capture
the local context related
to a classifier''s reliability.

We the

n present our main methodology STRIVE. STRIVE employs a
metaclassificat

ion approach to learn an improved model which
varies the combination rul

e by considering the local
reliability of the base classifiers via the i

ndicators. The
resulting models empirically outperform state-of-the-art

metaclassification approaches that do not use locality.
Next we an

alyze the contributions of the various reliability
indicators to the com

bination model and suggest informative
features to consider when redesig

ning the base classifiers.
Additionally we show how inductive transfer

methods can be
extended to increase the amount of labeled training data

for
learning a combination model by collapsing data traditionally
vie

wed as coming from different learning tasks.

Finally the combinatio

n approaches discussed are broadly
applicable to classification problems
other than topic
classification and we emphasize this with experiments
that
demonstrate STRIVE improves performance of action-item
detector

s in e-mail -- a task where both the semantics and base
classifier perfo

rmance are significantly different than topic
classification.

Spe

aker Bio:
Paul Bennett is currently a Postdoctoral Fellow in the Languag

e
Technologies Institute at Carnegie Mellon University where he
serve

s as Chief Learning Architect on the RADAR project.
Paul''s primary rese

arch interests are in text classification
information retrieval ensemb

le methods and calibration with
wider interests in statistical learnin

g and applications of
artificial intelligence in adaptive systems in gen

eral. His
published work includes research on classifier combination action-item detection calibration inductive transfer machine
transl

ation and recommender systems.
Paul received his Ph.D. (2006) from the
Computer Science
Department at Carnegie Mellon University.