ActUp: Analyzing and Consolidating tSNE and UMAP

Andrew Draganov, Jakob Rødsgaard Jørgensen, Katrine Scheel Nellemann, Davide Mottin, Ira Assent, Tyrus Berry, Çigdem Aslay

IJCAI 2023

pdf | code

A single method (GDR) can recreate tSNE and UMAP outputs just by changing the normalization.

A single method (GDR) can recreate tSNE and UMAP outputs just by changing the normalization.

TL;DR

One method (GDR) recreate both tSNE and UMAP embedding, since the difference between them is only the normalization.

In this paper:

  • We perform the first comprehensive analysis of the dif- ferences between tSNE and UMAP, showing the effect of each algorithmic choice on the embeddings.
  • We theoretically and experimentally show that changing the normalization is a sufficient condition for switching between the two methods.
  • We release simple, plug-and-play implementations of GDR, tSNE and UMAP that can toggle all of the iden- tified hyperparameters. Furthermore, GDR obtains em- beddings for both algorithms faster than UMAP.