jmlr-announce: Designing Committees of Models through Deliberate Weighting of Data Points

[David 'Pablo' Cohn]

The Journal of Machine Learning Research (www.jmlr.org) is pleased to 
announce publication of  the third paper in Volume 4:

--------------------------
Designing Committees of Models through Deliberate Weighting of Data Points
Stefan W. Christensen, Ian Sinclair and Philippa A. S. Reed
JMLR 4(Apr):39-66, 2003.

Abstract

In the adaptive derivation of mathematical models from data, each data 
point should contribute with a weight reflecting the amount of confidence 
one has in it. When no additional information for data confidence is 
available, all the data points should be considered equal, and are also 
generally given the same weight. In the formation of committees of models, 
however, this is often not the case and the data points may exercise 
unequal, even random, influence over the committee formation.

In this paper, a principled approach to committee design is presented. The 
construction of a committee design matrix is detailed through which each 
data point will contribute to the committee formation with a fixed weight, 
while contributing with different individual weights to the derivation of 
the different constituent models, thus encouraging model diversity whilst 
not biasing the committee inadvertently towards any particular data points. 
Not distinctly an algorithm, it is instead a framework within which several 
different committee approaches may be realised.

Whereas the focus in the paper lies entirely on regression, the principles 
discussed extend readily to classification.

----------------------------------------------------------------------------
This paper, and all previous papers, are available electronically at 
http://www.jmlr.org in PostScript and PDF formats. The papers of Volumes 1, 
2 and 3 are also available electronically from the JMLR website, and in 
hardcopy from the MIT Press; please see http://mitpress.mit.edu/JMLR for 
details.

-David Cohn, <david.cohn at acm.org>