[NRG] Reminder: NRG Meeting: Targeted Matrix Completion (Natali Ruchansky) @ Mon Dec 9, 2013 11am - 12pm (NRG Calendar)

Google Calendar calendar-notification at google.com
Sun Dec 8 10:59:56 EST 2013

This is a reminder for:

Title: NRG Meeting: Targeted Matrix Completion (Natali Ruchansky)
Title: Targeted Matrix Completion

Presenter: Natali Ruchansky

In many problems the input consists of a partially-observed matrix and the  
goal is to identify a low-rank matrix that best fits these observations.   
This task, known as matrix completion, has in recent years become  
recognized as fundamental in data analysis.   Traditional statistical  
approaches formulate the matrix completion problem as an optimization one,  
assuming that the set of observed entries is sufficient for accurate  
reconstruction with high probability.  A major drawback is that such  
methods give no indication of which entries have been accurately recovered  
and which not -- since they output some completion for any set of  
observations, even the smallest.
On the other hand, the recent structural completion methods (such as the  
Information Cascade Matrix Completion method we adopt) evaluate the  
positions of observed entries and indicate precisely which missing entries  
can be recovered accurately.  Here, we ask the following natural question:  
when the observations are not enough to recover all entries, how can we add  
observations to enable full completion?  One strategy is to randomly reveal  
entries until a density at which statistical methods can output a good  
estimate.  Unfortunately this density can be rather high, and so we wonder  
if there is a smarter way.  Through a deep understanding of structural  
completion, we introduce an algorithm called Order&Extend that chooses  
entries strategically.  Together with structural completion, our algorithm  
can identify exactly which entries can be recovered given the observation,  
and then specify how much and which additional information is needed for  
full completion.  We show that Order&Extend requires less (almost minimal)  
information and leads to better recovery than statistical methods.

Short Bio:
Natali received a Bachelor of Art from Boston University in Mathematics and  
Computer Science.  She is now a PhD student in the Data Management Lab at  
Boston University, working with Evimaria Terzi and Mark Crovella.  Her  
research interests span many topics including algorithmic data mining and  

When: Mon Dec 9, 2013 11am – 12pm Eastern Time
Where: MCS-148, 111 Cummington Mall, Boston, MA 02215
Calendar: NRG Calendar
     * Larissa Spinelli - creator

Event details:  

Invitation from Google Calendar: https://www.google.com/calendar/

You are receiving this email at the account nrg-l at cs.bu.edu because you are  
subscribed for reminders on calendar NRG Calendar.

To stop receiving these notifications, please log in to  
https://www.google.com/calendar/ and change your notification settings for  
this calendar.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://cs-mailman.bu.edu/pipermail/nrg-l/attachments/20131208/f3edd4c5/attachment.html>

More information about the NRG-L mailing list