Efficient Clustering for Stretched Mixtures: Landscape and Optimality

• Activity Number: 347 - Recent Advances in Clustering and Mixture Models Analysis
• Type: Topic-Contributed
• Date/Time: Thursday, August 12, 2021 : 10:00 AM to 11:50 AM
• Sponsor: Section for Statistical Programmers and Analysts
• Author(s): Kaizheng Wang* and Yuling Yan and Mateo Diaz
• Companies: Columbia University and Princeton University and Cornell University
• Keywords: clustering; dimension reduction; unsupervised learning; landscape; non-convex optimization

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 landscape properties as long as 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 even without good initialization. We also propose a general methodology for multi-class clustering tasks with flexible choices of feature transforms and loss objectives.