[Ml-stat-talks] NMF, LDA, latent variables seminar 3/14/13
jstorey at Princeton.EDU
Tue Mar 12 10:03:00 EDT 2013
Non-negative matrix factorization, latent Dirichlet allocation, etc...
> From: Valerie Marino <vmarino at math.princeton.edu>
> Subject: IDeAS Seminar for 3/14/13
> Date: March 12, 2013 9:49:30 AM EDT
> To: undisclosed-recipients:;
> Seminar: IDeAS Seminar
> Date: 3/14/13
> Time & Place: 102A McDonnell Hall @ 3:45 pm
> Speaker: Ankur Moitra - Institute for Advanced Study
> Title: Provable Algorithms for Nonnegative Matrix Factorization and Beyond
> The nonnegative matrix factorization problem has important applications throughout machine learning where it is used to uncover latent statistical relationships present in data, that can then be used for clustering, information retrieval, recommendation systems etc. As is often the case, this problem is NP-hard when considered in full generality. However, we introduce a sub-case called separable nonnegative matrix factorization that we believe is the right notion in various contexts. We give a polynomial time algorithm for this problem, and leverage this algorithm to efficiently learn the topics in a Latent Dirichlet Allocation model and various other topic models. In fact, these algorithms are not only interesting from a theoretical standpoint but in fact run orders of magnitude faster than the existing best algorithms for these tasks, without sacrificing and in many cases improving the quality of the output. There are many natural questions about how these approaches can be extended to more general settings, and whether these algorithms can be successfully applied to even larger data sets than we have had the tools to explore thus far.
> Valerie Marino
> Program Secretary
> Program in Applied & Computational Mathematics
> Tel: 609-258-3703
> Fax: 609-258-1735
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Ml-stat-talks