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:Machine Learning @ CUED
SUMMARY:Matrix Factorization and Relational Learning - Aji
t Paul Singh (CMU)
DTSTART;TZID=Europe/London:20080909T140000
DTEND;TZID=Europe/London:20080909T150000
UID:TALK13396AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/13396
DESCRIPTION:Matrix factorization is one of the workhorse metho
ds in data mining\, machine learning\, and informa
tion retrieval. We present a unified view of matri
x factorization models\, which includes weighted s
ingular value decompositions\, non-negative matrix
factorization\, probabilistic latent semantic ind
exing\, max-margin matrix factorization\, matrix c
o-clustering\, and generalizations of these models
to exponential family distributions. This unified
view leads to a class of optimization algorithms\
, based on alternating projections and stochastic
approximations\, which are well-suited to models o
f large\, sparse matrices.\n\nExtending upon our u
nified view of matrix factorization\, many types o
f relational data can be presented as a set of rel
ated matrices\, where shared dimensions correspond
to shared factors in a low-rank representation. W
e extend Bregman matrix factorization to a set of
related matrices\, illustrating the use of relatio
nal learning on a collaborative filtering problem.
\n\nThis talk is based primarily on two publicatio
ns: _Relational Learning via Collective Matrix Fac
torization_ (Singh & Gordon\, KDD-2008)\, and _A U
nified View of Matrix Factorization Models_ (Singh
& Gordon\, ECML/PKDD-2008).\n\n
LOCATION:Engineering Department\, CBL Room 438
CONTACT:Zoubin Ghahramani
END:VEVENT
END:VCALENDAR