I'm relatively new to t-SNE and it seems that unlike in other dimensionality-reduction techniques, the dimensions in t-SNE are hard to interpret. The contribution of the variables can easily be accessed in PCA and MCA, and one can regress on the dimensions of MDS to know how each variable affects the position of the points in the n-dimensional space. Since there is no linear relationship between the dimensions and the points in t-SNE (at least that's what I've read somewhere), isn't there a way to get to know how the variables explain the dimensions?

I motivate my choice of t-SNE because it takes much less time to represent the distances than e.g. MDS. I'm using R version 3.4.4 (64-bit) and the Rtsne function from the package of the same name on a packard bell (Intel(R) Celeron(R) CPU B830 @ 1.80 GHz with 8,00 G RAM).

Thanking you in advance.

  • $\begingroup$ The dimensions and distances are in general not interpretable, check out: distill.pub/2016/misread-tsne $\endgroup$
    – dcl
    May 23, 2018 at 0:40
  • $\begingroup$ MDS is also nonlinear. "one can regress on the dimensions of MDS" -- I think it does not make sense, but if you are fine with doing that, then you can as well do it for t-SNE. $\endgroup$
    – amoeba
    May 24, 2018 at 7:49


Browse other questions tagged or ask your own question.