Efficient Clustering for Stretched Mixtures: Landscape and Optimality
Kaizheng Wang, Yuling Yan, Mateo Diaz
Abstract
This paper considers a canonical clustering problem where one receives unlabeled samples drawn from a balanced mixture of two elliptical distributions and aims for a classifier to estimate the labels. Many popular methods including PCA and k-means require individual components of the mixture to be somewhat spherical, and perform poorly when they are stretched. To overcome this issue, we propose a non-convex program seeking for an affine transform to turn the data into a one-dimensional point cloud concentrating around -1 and 1, after which clustering becomes easy. Our theoretical contributions are two-fold: (1) we show that the non-convex loss function exhibits desirable geometric properties when the sample size exceeds some constant multiple of the dimension, and (2) we leverage this to prove that an efficient first-order algorithm achieves near-optimal statistical precision without good initialization. We also propose a general methodology for clustering with flexible choices of feature transforms and loss objectives.
BibTeX
@inproceedings{NEURIPS2020_f40ee694,
author = {Wang, Kaizheng and Yan, Yuling and Diaz, Mateo},
booktitle = {Advances in Neural Information Processing Systems},
editor = {H. Larochelle and M. Ranzato and R. Hadsell and M.F. Balcan and H. Lin},
pages = {21309--21320},
publisher = {Curran Associates, Inc.},
title = {Efficient Clustering for Stretched Mixtures: Landscape and Optimality},
url = {https://proceedings.neurips.cc/paper_files/paper/2020/file/f40ee694989b3e2161be989e7b9907fc-Paper.pdf},
volume = {33},
year = {2020}
}