Institute Homepage
Institute Homepage DE Sign In

Back

Empirical Inference Conference Paper 2007

A Dependence Maximization View of Clustering

no image
Empirical Inference
Arthur Gretton

We propose a family of clustering algorithms based on the maximization of dependence between the input variables and their cluster labels, as expressed by the Hilbert-Schmidt Independence Criterion (HSIC). Under this framework, we unify the geometric, spectral, and statistical dependence views of clustering, and subsume many existing algorithms as special cases (e.g. k-means and spectral clustering). Distinctive to our framework is that kernels can also be applied on the labels, which can endow them with particular structures. We also obtain a perturbation bound on the change in k-means clustering.

Author(s): Song, L. and Smola, AJ. and Gretton, A. and Borgwardt, KM.
Journal: Proceedings of the 24th Annual International Conference on Machine Learning (ICML 2007)
Pages: 815-822
Year: 2007
Month: June
Day: 0
Editors: Ghahramani, Z.
Publisher: ACM Press
Bibtex Type: Conference Paper (inproceedings)
Address: New York, NY, USA
DOI: 10.1145/1273496.1273599
Event Name: Twenty-Fourth Annual International Conference on Machine Learning (ICML 2007)
Event Place: Corvallis, OR, USA
Digital: 0
Electronic Archiving: grant_archive
ISBN: 978-1-59593-793-3
Language: en
Organization: Max-Planck-Gesellschaft
School: Biologische Kybernetik
Links:

BibTex 

@inproceedings{4471,
  title = {A Dependence Maximization View of Clustering},
  journal = {Proceedings of the 24th Annual International Conference on Machine Learning (ICML 2007)},
  abstract = {We propose a family of clustering algorithms based on the maximization of dependence between the input variables and their cluster labels, as expressed by the Hilbert-Schmidt Independence Criterion (HSIC). Under this framework, we unify the geometric, spectral, and statistical dependence views of clustering, and subsume many existing algorithms as special cases (e.g. k-means and spectral clustering). Distinctive to our framework is that kernels can also be applied on the labels, which can endow them with particular structures. We also obtain a perturbation bound on the change in k-means clustering.},
  pages = {815-822},
  editors = {Ghahramani, Z. },
  publisher = {ACM Press},
  organization = {Max-Planck-Gesellschaft},
  school = {Biologische Kybernetik},
  address = {New York, NY, USA},
  month = jun,
  year = {2007},
  slug = {4471},
  author = {Song, L. and Smola, AJ. and Gretton, A. and Borgwardt, KM.},
  month_numeric = {6}
}