Does there exist an implementation (EM-clustering) that allows you to specify the initial cluster positions beforehand? The datasets relevant for me are two dimensional and since just assign random cluster positions at start yield less accuracy than a clever initial setup I want to analyze 2d plots and specify my own initialization.
$\begingroup$
$\endgroup$
1
-
$\begingroup$ Yes, e.g., in ELKI you can provide initial cluster centers from Java. I'm fairly certain this will also be possible in sklearn and R. Just spend a few minutes with Google and looking at the codes. And consider contributing missing functionality to open source projects. $\endgroup$– Has QUIT--Anony-MousseDec 18, 2017 at 16:29
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
Maybe the most popular approach is to initialize GMM is through k-means, and use the result as the centers for the Gaussians. There are other approaches, but in practice that is the most common approach I have seen and used.
-
$\begingroup$ Actually the most popular, c.f. Mclust in R, seems to be to use hierarchical clustering. $\endgroup$ Dec 18, 2017 at 16:30