---------------------------------------
Statistics seminar
School of Mathematics
The University of Edinburgh
---------------------------------------
Friday 29st March, 3.10 pm Room 5323, JCMB
Speaker: Ludger Evers,
University of Glasgow
Geometrically-inspired structured priors for Bayesian partition-based models
Abstract
Partition-based regression models, which include both
recursive-partitioning methods such as CARTs as well as methods based on
tessellations, have been used successfully for more than a decade.
However, partition-based models typically lead to discontinuous
predictions which increases the variance of the predictions. Bayesian
approaches suffer in addition from rather poorly mixing MCMC algorithms.
The aim is to show how geometrically inspired structured joint priors on
the models inside each region can overcome both problems. The core idea
behind the construction of the prior is to assume a linear, rather than
constant, model inside each region and to penalise large differences
between neighbouring regions in the partition implied by the method.
This choice of prior reflects the geometry of the partition rather than
the process used to generate it. This is the core difference to related
methods suggested in the literature. The priors yield both “smooth-ish”
predictive surfaces and better mixing MCMC chains and thus yield a
powerful flexible regression method.
This seminar is joint with BioSS and is a part of Maxwell Institute
Statistics seminar.
There will be tea and coffee in the Mathematics Common Room (5212) after
the seminar.
The seminar websites:
http://www.maths.ed.ac.uk/events/statistics (in google calendar)
http://www.maths.ed.ac.uk/~nbochkin/StatisticsSeminar.html
--
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
You may leave the list at any time by sending the command
SIGNOFF allstat
to [log in to unmask], leaving the subject line blank.
|