I am using SOM for dimension reduction and visualization purpose (to put the same observations together). I am using kohonen r-package for the same.


For experimental purpose I took a 2-dimensional data (original):

  1. Applied hierarchical clustering on original data.
  2. Applied SOM on original data and then applied hierarchical clustering on top of that.

I got similar results.

So my question is: Is preserving the topological distances only advantage behind using SOM over clustering?

If yes, then why this is important? If no, what are other advantages of SOM over clustering?


You took a much too simple example.

If your input data is 2d, you don't need to use a SOM at all.

The purpose of a SOM is to put a 2d map over your data, even if that data is of higher dimensionality.

Apart from that, your question does not make much sense. You are comparing apples and oranges (SOM cannot be "better than clustering" because it solves a different problem). Furthermore, who says that SOMs are supposed to be better than anything?

Better comparisons may be SOM vs. tSNE, for example.

| cite | improve this answer | |
  • $\begingroup$ I completely agree, it is a bad example. No need to use SOM for 2D data. As the SOM also used for clustering/segmentation, so I think I am comparing apples vs apples. By clustering I meant clustering techniques. $\endgroup$ – Artiga Jun 8 '18 at 11:27
  • 1
    $\begingroup$ SOM itself doesn't do clustering. You can use kmeans somewhat on the mapped data. But that isn't very sound or reliable, do is ask but popular to use SOMs with clustering. It's more of a visualization thing. $\endgroup$ – Has QUIT--Anony-Mousse Jun 9 '18 at 23:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.