Saved in:
Bibliographic Details
Main Authors: Sun, Qiang, Zhang, Archer Gong, Liu, Chenyu, Tan, Kean Ming
Format: Preprint
Published: 2019
Subjects:
Online Access:https://arxiv.org/abs/1906.09581
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909343177244672
author Sun, Qiang
Zhang, Archer Gong
Liu, Chenyu
Tan, Kean Ming
author_facet Sun, Qiang
Zhang, Archer Gong
Liu, Chenyu
Tan, Kean Ming
contents Convex clustering is a convex relaxation of the $k$-means and hierarchical clustering. It involves solving a convex optimization problem with the objective function being a squared error loss plus a fusion penalty that encourages the estimated centroids for observations in the same cluster to be identical. However, when data are contaminated, convex clustering with a squared error loss fails even when there is only one arbitrary outlier. To address this challenge, we propose a resistant convex clustering method. Theoretically, we show that the new estimator is resistant to arbitrary outliers: it does not break down until more than half of the observations are arbitrary outliers. Perhaps surprisingly, the fusion penalty can help enhance resistance by fusing the estimators to the cluster centers of uncontaminated samples, but not the other way around. Numerical studies demonstrate the competitive performance of the proposed method.
format Preprint
id arxiv_https___arxiv_org_abs_1906_09581
institution arXiv
publishDate 2019
record_format arxiv
spellingShingle Resistant convex clustering: How does the fusion penalty enhance resistantance?
Sun, Qiang
Zhang, Archer Gong
Liu, Chenyu
Tan, Kean Ming
Methodology
Convex clustering is a convex relaxation of the $k$-means and hierarchical clustering. It involves solving a convex optimization problem with the objective function being a squared error loss plus a fusion penalty that encourages the estimated centroids for observations in the same cluster to be identical. However, when data are contaminated, convex clustering with a squared error loss fails even when there is only one arbitrary outlier. To address this challenge, we propose a resistant convex clustering method. Theoretically, we show that the new estimator is resistant to arbitrary outliers: it does not break down until more than half of the observations are arbitrary outliers. Perhaps surprisingly, the fusion penalty can help enhance resistance by fusing the estimators to the cluster centers of uncontaminated samples, but not the other way around. Numerical studies demonstrate the competitive performance of the proposed method.
title Resistant convex clustering: How does the fusion penalty enhance resistantance?
topic Methodology
url https://arxiv.org/abs/1906.09581