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:Learning Markov Networks for Mixed Big Data: Appli
cations to Cancer Genomics - Genevera Allen\, Rice
University
DTSTART;TZID=Europe/London:20150305T160000
DTEND;TZID=Europe/London:20150305T170000
UID:TALK57846AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/57846
DESCRIPTION:"Mixed Data'' comprising a large number of heterog
eneous variables (e.g. count\, binary\,\ncontinuo
us\, skewed continuous\, among other data types) i
s prevalent in varied areas such as\nimaging genet
ics\, national security\, social networking\, Inte
rnet advertising\, and our particular\nmotivation
- high-throughput integrative genomics. There hav
e been limited efforts at\nstatistically modeling
such mixed data jointly\, in part because of the l
ack of computationally\namenable multivariate dist
ributions that can capture direct dependencies bet
ween variables of\ndifferent types. \nIn this tal
k\, we address this by introducing several new cla
sses of Markov Random Fields (MRFs)\,\nor graphica
l models\, that yield joint densities over mixed v
ariables. To begin\, we present a\nnovel class of
MRFs arising when all node-conditional distributio
ns follow univariate\nexponential family distribut
ions that\, for instance\, yield novel Poisson gra
phical models. \nNext\, we introduce extensions of
this for Mixed MRF distributions. Unfortunately\
, these\nformulations can place severe and unreali
stic restrictions on the parameter space. To reme
dy\nthis\, we we introduce a class of mixed condit
ional random field distributions\, that are then\n
chained according to a block-directed acyclic grap
h to form a new class of so-called Block\nDirected
Markov Random Fields (BDMRFs). The Markov indepen
dence graph structure underlying our\nBDMRF then h
as both directed and undirected edges. \n\nWe will
briefly review the theoretical properties of thes
e models and introduce penalized\nconditional like
lihood estimators with statistical guarantees for
learning the underlying mixed\nnetwork structure.
Simulations as well as an application to integrati
ve cancer genomics\ndemonstrate the versatility of
our methods. In our particular example\, we lear
n integrative\ngenomic networks from breast cancer
next generation sequencing expression data and mu
tation data\nthat yield several interesting findin
gs.\n \nJoint work with Eunho Yang\, Pradeep Raviu
kmar\, Zhandong Liu\, Yulia Baker\, and Ying-Wooi
Wan.
LOCATION:MR12\, Centre for Mathematical Sciences\, Wilberf
orce Road\, Cambridge
CONTACT:
END:VEVENT
END:VCALENDAR