Interpreting t-SNE dimensions in terms of the original features

Question

I have a dataset of 223 multilabels. I'm sure some of those labels are correlated so I want to merge them. For that I'm using sklearn tSNE to reduce the dimensions.

How can I "describe" the new dimensions in terms of the former ones, I need something like this:

0.5(rock) + 0.1(pop)       generated popRock    
0.2(rock) + 0.5(christian) generated christianRock

score 3 · Answer 1 · answered Dec 01 '18 at 20:24

3

This isn't possible. At a high level you can think of tSNE as pulling similar points together and pushing away dissimilar points with simulated elastic springs. There is no linear correspondence between old and new dimensions, and it's not even possible to compute a nonlinear mapping from old to new dimensions.

answered Dec 01 '18 at 20:24

shimao

22,706
2
42
81

Thanks, even with other dimmension reduction methods? – Hohenheimsenberg Dec 02 '18 at 00:20
Nevermind I found an answer https://stats.stackexchange.com/questions/229092/how-to-reverse-pca-and-reconstruct-original-variables-from-several-principal-com – Hohenheimsenberg Dec 02 '18 at 00:39

Interpreting t-SNE dimensions in terms of the original features

1 Answers1