MPI-INF D5 Publications: Report

Technical, Research Report
@TechReport
Technischer-, Forschungsbericht

Show entries of:

this year (2019) | last year (2018) | two years ago (2017) | Notes URL

Action:

Options:

Library locked

Goto entry point

Author, Editor

Author(s):

Gemulla, Rainer
Haas, Peter J.
Nijkamp, Erik
Sismanis, Yannis

dblp
dblp
dblp
dblp

Not MPG Author(s):

Haas, Peter J.
Nijkamp, Erik
Sismanis, Yannis

Editor(s):

BibTeX Citekey*:	gemulla11
Language:	English

Title, Institution

Title*:	Large-scale matrix factorization with distributed stochastic gradient descent
Institution*:	IBM Almaden Research Center
Publishers or Institutions Address*:	San Jose, CA
Type:	Technical Report

No, Year, pp.,

Number*:	RJ10481	Pages*:	47
Month:	March	VG Wort Pages*:
Year*:	2011	ISBN/ISSN:

DOI:

Note, Abstract, ©

Note:
(LaTeX) Abstract:	As Web 2.0 and enterprise-cloud applications have proliferated, data mining algorithms increasingly need to be (re)designed to handle web-scale datasets. For this reason, low-rank matrix factorization has received a lot of attention in recent years, since it is fundamental to a variety of mining tasks, such as topic detection and collaborative filtering, that are increasingly being applied to massive datasets. We provide a novel algorithm to approximately factor large matrices with millions of rows, millions of columns, and billions of nonzero elements. Our approach rests on stochastic gradient descent (SGD), an iterative stochastic optimization algorithm; the idea is to exploit the special structure of the matrix factorization problem to develop a new ``stratified'' SGD variant that can be fully distributed and run on web-scale datasets using, e.g., MapReduce. The resulting distributed SGD factorization algorithm, called DSGD, provides good speed-up and handles a wide variety of matrix factorizations. We establish convergence properties of DSGD using results from stochastic approximation theory and regenerative process theory, and also describe the practical techniques used to optimize performance in our DSGD implementation. Experiments suggest that DSGD converges significantly faster and has better scalability properties than alternative algorithms.
Categories / Keywords:
Copyright Message:
HyperLinks / References / URLs:	http://www.almaden.ibm.com/cs/people/peterh/dsgdTechRep.pdf
Personal Comments:
File Upload:
Download Access Level:	Public

Correlation

MPG Unit:	Max-Planck-Institut für Informatik

MPG Subunit:	Databases and Information Systems Group
Appearance:	MPII WWW Server, MPII FTP Server, MPG publications list, university publications list, working group publication list, Fachbeirat, VG Wort

BibTeX Entry:

@TECHREPORT{gemulla11,
AUTHOR = {Gemulla, Rainer and Haas, Peter J. and Nijkamp, Erik and Sismanis, Yannis},
TITLE = {Large-scale matrix factorization with distributed stochastic gradient descent},
YEAR = {2011},
TYPE = {Technical Report},
INSTITUTION = {IBM Almaden Research Center},
NUMBER = {RJ10481},
PAGES = {47},
ADDRESS = {San Jose, CA},
MONTH = {March},
}

Entry last modified by Anja Becker, 03/22/2012

Edit History (please click the blue arrow to see the details)

	Editor(s) [Library]	Created 03/08/2011 01:30:17 PM
Revisions 6. 5. 4. 3. 2.	Editor(s) Anja Becker Anja Becker Anja Becker Rainer Gemulla Rainer Gemulla	Edit Dates 22.03.2012 14:19:15 15.03.2012 16:28:44 12.03.2012 11:12:35 01/08/2012 01:16:40 PM 03/08/2011 01:46:05 PM

Attachment Section

Attachment Section

View attachments here:

dsgdTechRep.pdf

dsgdTechRep.pdf