Why it is recommended to use T SNE to reduce to 2-3 dims and not higher dim?

Question

• According to wiki it is recommenced to use T-SNE to map to 2-3 dimensional.
• I can understand this , if we want to visualizing the data.
• If we want to reduce the number of features (i.e from 30 features to 5 dims), is it recommended to do this with T-SNE ? or we should use other dimensional reduction algorithm ?

Answer 1

Big Alarm!

T-SNE is NOT a dimensionality reduction algorithm (like PCA, LLE, UMAP, etc.). It is ONLY for visualization, and for that sake, more than 3 dimensions does not make sense.

T-SNE is not a parametric method so you do not get abase vector representation based on which you reduce dimensionality of a new dataset (validation, test). Thats why it can not be used for dimensionality reduction.

It is calculated stochastically only based on the data it sees, so if you use it on the train set, there is no way to do the same calculation for your test set thus no modeling with T-SNE.

If you see Sklearn functions, for PCA and other dimensionality algorithms you see both fit(), transform() and fit_transform() functions but for T-SNE you have only fit() and fit_transform() because you will have no model to only transform() a new dataset.

I tried to be intuitive. If you need more technical explanation just drop a comment.