Spectral Clustering via the Power Method - Provably

Christos Boutsidis, Prabhanjan Kambadur, Alex Gittens
; Proceedings of the 32nd International Conference on Machine Learning, PMLR 37:40-48, 2015.

Abstract

Spectral clustering is one of the most important algorithms in data mining and machine intelligence; however, its computational complexity limits its application to truly large scale data analysis. The computational bottleneck in spectral clustering is computing a few of the top eigenvectors of the (normalized) Laplacian matrix corresponding to the graph representing the data to be clustered. One way to speed up the computation of these eigenvectors is to use the “power method” from the numerical linear algebra literature. Although the power method has been empirically used to speed up spectral clustering, the theory behind this approach, to the best of our knowledge, remains unexplored. This paper provides the first such rigorous theoretical justification, arguing that a small number of power iterations suffices to obtain near-optimal partitionings using the approximate eigenvectors. Specifically, we prove that solving the k-means clustering problem on the approximate eigenvectors obtained via the power method gives an additive-error approximation to solving the k-means problem on the optimal eigenvectors.

Cite this Paper


BibTeX
@InProceedings{pmlr-v37-boutsidis15, title = {Spectral Clustering via the Power Method - Provably}, author = {Christos Boutsidis and Prabhanjan Kambadur and Alex Gittens}, pages = {40--48}, year = {2015}, editor = {Francis Bach and David Blei}, volume = {37}, series = {Proceedings of Machine Learning Research}, address = {Lille, France}, month = {07--09 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v37/boutsidis15.pdf}, url = {http://proceedings.mlr.press/v37/boutsidis15.html}, abstract = {Spectral clustering is one of the most important algorithms in data mining and machine intelligence; however, its computational complexity limits its application to truly large scale data analysis. The computational bottleneck in spectral clustering is computing a few of the top eigenvectors of the (normalized) Laplacian matrix corresponding to the graph representing the data to be clustered. One way to speed up the computation of these eigenvectors is to use the “power method” from the numerical linear algebra literature. Although the power method has been empirically used to speed up spectral clustering, the theory behind this approach, to the best of our knowledge, remains unexplored. This paper provides the first such rigorous theoretical justification, arguing that a small number of power iterations suffices to obtain near-optimal partitionings using the approximate eigenvectors. Specifically, we prove that solving the k-means clustering problem on the approximate eigenvectors obtained via the power method gives an additive-error approximation to solving the k-means problem on the optimal eigenvectors.} }
Endnote
%0 Conference Paper %T Spectral Clustering via the Power Method - Provably %A Christos Boutsidis %A Prabhanjan Kambadur %A Alex Gittens %B Proceedings of the 32nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2015 %E Francis Bach %E David Blei %F pmlr-v37-boutsidis15 %I PMLR %J Proceedings of Machine Learning Research %P 40--48 %U http://proceedings.mlr.press %V 37 %W PMLR %X Spectral clustering is one of the most important algorithms in data mining and machine intelligence; however, its computational complexity limits its application to truly large scale data analysis. The computational bottleneck in spectral clustering is computing a few of the top eigenvectors of the (normalized) Laplacian matrix corresponding to the graph representing the data to be clustered. One way to speed up the computation of these eigenvectors is to use the “power method” from the numerical linear algebra literature. Although the power method has been empirically used to speed up spectral clustering, the theory behind this approach, to the best of our knowledge, remains unexplored. This paper provides the first such rigorous theoretical justification, arguing that a small number of power iterations suffices to obtain near-optimal partitionings using the approximate eigenvectors. Specifically, we prove that solving the k-means clustering problem on the approximate eigenvectors obtained via the power method gives an additive-error approximation to solving the k-means problem on the optimal eigenvectors.
RIS
TY - CPAPER TI - Spectral Clustering via the Power Method - Provably AU - Christos Boutsidis AU - Prabhanjan Kambadur AU - Alex Gittens BT - Proceedings of the 32nd International Conference on Machine Learning PY - 2015/06/01 DA - 2015/06/01 ED - Francis Bach ED - David Blei ID - pmlr-v37-boutsidis15 PB - PMLR SP - 40 DP - PMLR EP - 48 L1 - http://proceedings.mlr.press/v37/boutsidis15.pdf UR - http://proceedings.mlr.press/v37/boutsidis15.html AB - Spectral clustering is one of the most important algorithms in data mining and machine intelligence; however, its computational complexity limits its application to truly large scale data analysis. The computational bottleneck in spectral clustering is computing a few of the top eigenvectors of the (normalized) Laplacian matrix corresponding to the graph representing the data to be clustered. One way to speed up the computation of these eigenvectors is to use the “power method” from the numerical linear algebra literature. Although the power method has been empirically used to speed up spectral clustering, the theory behind this approach, to the best of our knowledge, remains unexplored. This paper provides the first such rigorous theoretical justification, arguing that a small number of power iterations suffices to obtain near-optimal partitionings using the approximate eigenvectors. Specifically, we prove that solving the k-means clustering problem on the approximate eigenvectors obtained via the power method gives an additive-error approximation to solving the k-means problem on the optimal eigenvectors. ER -
APA
Boutsidis, C., Kambadur, P. & Gittens, A.. (2015). Spectral Clustering via the Power Method - Provably. Proceedings of the 32nd International Conference on Machine Learning, in PMLR 37:40-48

Related Material