A Comparative Study of Local Search Algorithms for Correlation Clustering

Research output: Contribution to book/conference proceedings/anthology/reportConference contributionContributedpeer-review

Contributors

  • Evgeny Levinkov - , Max Planck Institute for Informatics (Author)
  • Alexander Kirillov - , Chair of Image Processing (Author)
  • Bjoern Andres - , Max Planck Institute for Informatics (Author)

Abstract

This paper empirically compares four local search algorithms for correlation clustering by applying these to a variety of instances of the correlation clustering problem for the tasks of image segmentation, hand-written digit classification and social network analysis. Although the local search algorithms establish neither lower bounds nor approximation certificates, they converge monotonously to a fixpoint, offering a feasible solution at any time. For some algorithms, the time of convergence is affordable for all instances we consider. This finding encourages a broader application of correlation clustering, especially in settings where the number of clusters is not known and needs to be estimated from data.

Details

Original languageEnglish
Title of host publicationPattern Recognition
EditorsVolker Roth, Thomas Vetter
PublisherSpringer, Cham
Pages103–114
ISBN (electronic)978-3-319-66709-6
ISBN (print)978-3-319-66708-9
Publication statusPublished - 2017
Peer-reviewedYes

Publication series

SeriesLecture Notes in Computer Science
Volume10496
ISSN0302-9743

External IDs

ORCID /0000-0001-5036-9162/work/161888479