Clustering with t-SNE, Provably

Кластеризация с помощью t-SNE: строгое обоснование
George C. Linderman, Stefan Steinerberger
2019-01-01

cluster recoveryearly exaggerationspectral clusteringswiss rollt-SNE
t-distributed stochastic neighborhood embedding (t-SNE), a clustering and visualization method proposed by van der Maaten and Hinton in 2008, has rapidly become a standard tool in a number of natural sciences. Despite its overwhelming success, there is a distinct lack of mathematical foundations, and the inner workings of the algorithm are not well understood. The purpose of this paper is to prove that t-SNE is able to recover well-separated clusters; more precisely, we prove that t-SNE in the “early exaggeration” phase, an optimization technique proposed by van der Maaten and Hinton [ J. Mach. Learn. Res., 9 (2008), pp. 2579--2605] and van der Maaten [ J. Mach. Learn. Res., 15 (2014), pp. 3221--3245], can be rigorously analyzed. As a byproduct, the proof suggests novel ways for setting the exaggeration parameter $\alpha$ and step size $h$. Numerical examples illustrate the effectiveness of these rules: in particular, the quality of embedding of topological structures (e.g., the swiss roll) improves. We also discuss a connection to spectral clustering methods.
1
Numerical examples show that the proposed parameter-setting rules improve embeddings of topological structures such as the swiss roll.
2
The analysis establishes conditions under which early-exaggeration t-SNE can separate well-separated clusters.
3
The paper identifies a connection between early-exaggeration t-SNE and spectral clustering methods.
4
The paper provides a rigorous mathematical analysis of t-SNE’s early exaggeration phase for recovering well-separated clusters.
5
The proof suggests novel rules for selecting the exaggeration parameter α and optimization step size h.

t-SNE in the early exaggeration phase applied to well-separated clusters and topological data structures

Theoretical cluster-recovery guarantees and embedding behavior of t-SNE, including the effects of the exaggeration parameter and step size

Publication Details
Publication Date
2019-01-01
Journal
Publisher
ISSN
Cited by
315
Access Type
Author Information
Authors
George C. Linderman
Stefan Steinerberger
Explore further
Open the scid.ai AI chat with a ready-made request: it will find papers on a similar topic and help build a literature review.
Find similar papers in the chat →
Make a presentation
100%