TIME: 1:30-2:20 pm,  April 10, 2007


SPEAKER: Anthony Wirth
         University of Melbourne

TITLE: Improvements in consensus clustering and psychometric function 


In this talk I will present the results of two projects: one on designing
practical consensus clustering algorithms from approximation techniques, 
other on improving psychometric function estimation. Both of these
are outcomes of rewarding collaborations with a niche group of
research-focused undergraduates at my institution.

Consensus clustering is essentially the problem of finding the
1-median in the space of clusterings, provided there is a sensible
metric on clusterings of the same data. Until Ailon et al's work a few
years ago the trivial "pick-the-best-input-clustering" method was the
only known approximation algorithm [factor 2]. Building on work of
Gionis et al, Michael Bertolacci and I have shown that sampling the
data, combined with good approximation and 'unsampling' methods yields
high-quality consensuses quickly.

Psychophysics is the study of how the response, or lack thereof,
varies when a subject is shown different kinds of stimuli, or more
often stimuli of varying intensities. The latter case is frequently
modelled with a psychometric function, which somewhat resembles a
cumulative distribution function. With Andrew Turpin (RMIT
University), Elena Kelareva and James Mewing, we showed how to make a
small improvement to an information-theory inspired method for
estimating the threshold of a psychometric function.

Tony Wirth was an undergraduate and masters student at the University
of Melbourne, majoring in both Computer Science and Statistics. His
PhD, completed in 2004, was with the theory group at Princeton
University. He has been on the faculty of the Department of Computer
Science and Software Engineering at Melbourne for just over two years.