Clustering by Left-Stochastic Matrix Factorization
Abstract
We propose clustering samples given their pairwise similarities by factorizing the similarity matrix into the product of a cluster probability matrix and its transpose. We propose a rotation-based algorithm to compute this left-stochastic decomposition (LSD). Theoretical results link the LSD clustering method to a soft kernel k-means clustering, give conditions for when the factorization and clustering are unique, and provide error bounds. Experimental results on simulated and real similarity datasets show that the proposed method reliably provides accurate clusterings.
Cite
Text
Arora et al. "Clustering by Left-Stochastic Matrix Factorization." International Conference on Machine Learning, 2011.Markdown
[Arora et al. "Clustering by Left-Stochastic Matrix Factorization." International Conference on Machine Learning, 2011.](https://mlanthology.org/icml/2011/arora2011icml-clustering/)BibTeX
@inproceedings{arora2011icml-clustering,
title = {{Clustering by Left-Stochastic Matrix Factorization}},
author = {Arora, Raman and Gupta, Maya R. and Kapila, Amol and Fazel, Maryam},
booktitle = {International Conference on Machine Learning},
year = {2011},
pages = {761-768},
url = {https://mlanthology.org/icml/2011/arora2011icml-clustering/}
}