BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Statistics
SUMMARY:Advances in Bayesian Latent Factor Modelling: The
Incredible Shrinking Model - Daniel Merl (Duke Uni
versity)
DTSTART;TZID=Europe/London:20080229T140000
DTEND;TZID=Europe/London:20080229T150000
UID:TALK10244AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/10244
DESCRIPTION:Recent advances in the methodology and application
of Bayesian latent factor models are discussed.
Sparse latent factor modelling\, in which the now-
standard latent factor modelling framework is coup
led with so-called sparsity\, or variable selectio
n\, prior distributions on regression parameters\,
has found significant application in the analysis
of data falling under the ``large p\, small n'' h
eading. This includes applications in genomics an
d finance\, for which a primary goal is to charact
erize variation in a very high-dimensional respons
e in terms of a concise set of factors with sparse
loadings. \nA useful consequence of sparsity in f
actor loadings is the increased opportunity for as
cribing specific scientific interpretation to the
latent factors. For example\, analysis of gene ex
pression microarrays in cancer studies has uncover
ed latent factors that have been demonstrated to b
e useful for predicting certain clinical outcomes.
The loadings associated with such factors therefo
re represent signatures\, in response-space\, of t
hose outcomes. \n\nTo the extent that signatures w
ith meaningful interpretations are thought to repr
esent underlying structural features of the system
under study\, such as groups of genes whose patte
rns of co-expression have causal relationships wit
h known phenotypes\, in many cases prior belief ab
out latent factor structure in new data should be
informed by posterior inferences drawn previously
on similar data. We extend the current class of s
parse latent factor models to include informative
variable selection priors on both latent factor lo
adings and latent factors. Through this approach\
, previously inferred signatures are projected ont
o and refined by new data. Examples from cancer g
enomics demonstrate how this sort of targeted late
nt factor search improves the scientific interpret
ability of the model\, and allows a direct query o
f the contribution of a known signature towards ex
plaining variation in new data. \n\nMore generally
\, this methodology defines a flexible class of fa
ctor models that through the model fitting process
``shrinks'' to the minimum number of covariates a
nd latent factors supported by the data. In order
to achieve total shrinkage\, we describe a new ty
pe of nonparametric variable selection prior for t
he variance components of the model. The prior inc
orporates a hierarchical Dirichlet process to indu
ce clustering of variance parameters within sample
groups\, while borrowing information across group
s to estimate mixture components. When a control g
roup is present\, the prior for non-control group
variances places additional probablility on a poin
t mass located at the control variance. In this w
ay\, variances are clustered both within and acros
s sample groups\, yielding a parsimonious represen
tation of the degree of heteroscedasticity in the
data. Hence\, the incredible shrinking model. \n\n
LOCATION:MR12\, CMS\, Wilberforce Road\, Cambridge\, CB3 0W
B
CONTACT:
END:VEVENT
END:VCALENDAR