BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Bayesian sequential design in matrix factorisation models - Sergio
  Bacallado (University of Cambridge)
DTSTART:20160725T110000Z
DTEND:20160725T113000Z
UID:TALK66838@talks.cam.ac.uk
CONTACT:INI IT
DESCRIPTION:<span>Co-author: Annie Marsden (University of Cambridge)  <br>
 </span> <span><br>Many problems in high-dimensional statistics rely on low
 -rank decompositions  of matrices. Examples include matrix completion\, re
 commender systems or  collaborative filtering\, and graph clustering or co
 mmunity detection. Most  commonly\, estimates are obtained by solving an o
 ptimisation problem through SDP  relaxations\, expectation maximisation\, 
 or projected gradient descent algorithms.  Bayesian analogs of these proce
 dures provide estimates of uncertainty\, but these  are rarely exploited i
 n practice. In this talk\, we explore how the posterior  distribution in m
 atrix factorisation models can be put to use in sequential  design problem
 s. Bayesian procedures such as Thompson sampling and the Bayesian  UCB hav
 e been shown to achieve optimal regret in Multi-Arm Bandit problems. We  p
 resent a simulation study supporting similar strategies in recommender sys
 tems.</span>
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR