1

Given the discussion on this answer what appropriate assumptions and contingencies would allow tSNE to have an informative interpretation?

It's hard to come away from that answer feeling like I can ever trust results from tSNE to ever show me if higher dimension data has meaningful separation. Are there other methods that do a better job?

Oberon Quinn
  • 41
  • 1
    My personal experience is that t-SNE is prone to introducing the impression of separation where separation in fact doesn't exist or is much weaker than it looks like. On the other hand, t-SNE can occasionally reveal separation in high-dimensional space that is otherwise hard to find. One thing you could do is to look at the original distances within and between "clusters" that you see in a t-SNE plot to see to what extent the view is distorted. – Christian Hennig Aug 14 '23 at 22:04

0 Answers0