A scalable approach to spectral clustering with SDD solvers

Nguyen Lu Dang Khoa, Sanjay Chawla

Research output: Contribution to journalArticle

6 Citations (Scopus)

Abstract

The promise of spectral clustering is that it can help detect complex shapes and intrinsic manifold structure in large and high dimensional spaces. The price for this promise is the expensive computational cost for computing the eigen-decomposition of the graph Laplacian matrix—so far a necessary subroutine for spectral clustering. In this paper we bypass the eigen-decomposition of the original Laplacian matrix by leveraging the recently introduced near-linear time solver for symmetric diagonally dominant (SDD) linear systems and random projection. Experiments on several synthetic and real datasets show that the proposed approach has better clustering quality and is faster than the state-of-the-art approximate spectral clustering methods.

Original languageEnglish
Pages (from-to)289-308
Number of pages20
JournalJournal of Intelligent Information Systems
Volume44
Issue number2
DOIs
Publication statusPublished - 2015
Externally publishedYes

Fingerprint

Decomposition
Subroutines
Linear systems
Costs
Experiments

Keywords

  • Random projection
  • Resistance distance
  • SDD solver
  • Spectral clustering

ASJC Scopus subject areas

  • Artificial Intelligence
  • Information Systems
  • Hardware and Architecture
  • Computer Networks and Communications
  • Software

Cite this

A scalable approach to spectral clustering with SDD solvers. / Khoa, Nguyen Lu Dang; Chawla, Sanjay.

In: Journal of Intelligent Information Systems, Vol. 44, No. 2, 2015, p. 289-308.

Research output: Contribution to journalArticle

@article{6945b387f90c41ad819c77b3f2a2f8e9,
title = "A scalable approach to spectral clustering with SDD solvers",
abstract = "The promise of spectral clustering is that it can help detect complex shapes and intrinsic manifold structure in large and high dimensional spaces. The price for this promise is the expensive computational cost for computing the eigen-decomposition of the graph Laplacian matrix—so far a necessary subroutine for spectral clustering. In this paper we bypass the eigen-decomposition of the original Laplacian matrix by leveraging the recently introduced near-linear time solver for symmetric diagonally dominant (SDD) linear systems and random projection. Experiments on several synthetic and real datasets show that the proposed approach has better clustering quality and is faster than the state-of-the-art approximate spectral clustering methods.",
keywords = "Random projection, Resistance distance, SDD solver, Spectral clustering",
author = "Khoa, {Nguyen Lu Dang} and Sanjay Chawla",
year = "2015",
doi = "10.1007/s10844-013-0285-0",
language = "English",
volume = "44",
pages = "289--308",
journal = "Journal of Intelligent Information Systems",
issn = "0925-9902",
publisher = "Springer Netherlands",
number = "2",

}

TY - JOUR

T1 - A scalable approach to spectral clustering with SDD solvers

AU - Khoa, Nguyen Lu Dang

AU - Chawla, Sanjay

PY - 2015

Y1 - 2015

N2 - The promise of spectral clustering is that it can help detect complex shapes and intrinsic manifold structure in large and high dimensional spaces. The price for this promise is the expensive computational cost for computing the eigen-decomposition of the graph Laplacian matrix—so far a necessary subroutine for spectral clustering. In this paper we bypass the eigen-decomposition of the original Laplacian matrix by leveraging the recently introduced near-linear time solver for symmetric diagonally dominant (SDD) linear systems and random projection. Experiments on several synthetic and real datasets show that the proposed approach has better clustering quality and is faster than the state-of-the-art approximate spectral clustering methods.

AB - The promise of spectral clustering is that it can help detect complex shapes and intrinsic manifold structure in large and high dimensional spaces. The price for this promise is the expensive computational cost for computing the eigen-decomposition of the graph Laplacian matrix—so far a necessary subroutine for spectral clustering. In this paper we bypass the eigen-decomposition of the original Laplacian matrix by leveraging the recently introduced near-linear time solver for symmetric diagonally dominant (SDD) linear systems and random projection. Experiments on several synthetic and real datasets show that the proposed approach has better clustering quality and is faster than the state-of-the-art approximate spectral clustering methods.

KW - Random projection

KW - Resistance distance

KW - SDD solver

KW - Spectral clustering

UR - http://www.scopus.com/inward/record.url?scp=84925289609&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84925289609&partnerID=8YFLogxK

U2 - 10.1007/s10844-013-0285-0

DO - 10.1007/s10844-013-0285-0

M3 - Article

VL - 44

SP - 289

EP - 308

JO - Journal of Intelligent Information Systems

JF - Journal of Intelligent Information Systems

SN - 0925-9902

IS - 2

ER -