Bayesian Classification Theory.pdf

Ba y esian Classi
cation Theory
T ec hnical Rep ort FIA
Robin Hanson John Stutz P eter Cheeseman

Sterling Soft w are NASA RIA CS
Arti
cial In telligence Researc h Branc h
NASA Ames Researc hCen ter
Mail Stop
Mo
et Field
CA
USA
Email
last
name
ptolem y
arc
nasa
go v
Auto class I I I
the most recen t released v ersion
com
Abstract
bines real and discrete data
allo ws some data to b e miss
The task of inferring a set of classes and class
ing
and automatically c ho oses the n um b er of classes
descriptions most lik ely to explain a giv en
from
rst principles
Extensiv e testing has indicated
data set can b e placed on a
rm theoretical
that it generally pro duces signi
can t and useful results
foundation using Ba y esian statistics
Within
the mo dels it uses
rather than
for example
inadequate
this framew ork
and using v arious mathemat
searc h heuristics
AutoClass I I I assumes that all at
ical and algorithmic appro ximations
the Au
tributes are relev an t
that they are indep enden tofeac h
toClass system searc hes for the most proba
other within eac h class
and that classes are m utually
ble classi
cations
automatically c ho osing the
exclusiv e
Recen t extensions
em b o died in Auto class IV
n um b er of classes plexit y of class de
let us relax t w o of these a