0
$\begingroup$

Consider we have two classes of points. Both of them come from a multivariate normal distribution with an unrestricted covariance matrix. Let's assume, that the densities of those distributions do not overlap much, i. e. $p_1(x) \cdot p_2(x) \approx 0$.

Will methods like TSNE, UMAP, PCA project those points (in a general case, since some of those methods are stochastic) into two visually separate clusters on a 2D plane?

I know this question is very broad, but I was wondering if there is a reason to search for clusters using gaussian mixture fitting or k-means in a high dimensional data, if TSNE projection shows no cluster structure. Or to put it even simpler: I know non-gaussian and non-oval clusters (embedded on some complex manifold) could exist if dimensionality reduction projection is a single blob of points, but I am not sure if this excludes the possibility of the clusters being multvariate normal.

Improve this question
$\endgroup$
2
  • $\begingroup$ Have you simulated it? $\endgroup$
    – Dave
    May 6 at 22:29
  • $\begingroup$ @Dave No I did not, as I don't think simulating things in spaces of, say, 50000 dimensions can be viable. But maybe there is a good method of generating such samples, but I know of none. $\endgroup$
    – Karol Szustakowski
    May 8 at 8:02

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.