Is it normal that SOM clusters the instances with the "versicolor" class into multiple different BMUs?

Question

I have trained (with different sizes, learning rates, and epochs) a SOM network to cluster the Iris dataset. The instances associated with the class setosa have been mainly fitted to a 1-2 BMUs. In the case of virginica, the instances have also be associated with only a few BMUs. However, in the case of versicolor instances, many BMUs have been associated with them.

Is this normal?

Now, I have a diagram. It doesn't look bad.

score 1 · Answer 1 · answered Sep 14 '18 at 12:07

This is not a bug, assuming you hve implemented the SOM properly with best decay rules of learning rate and neighbourhood strength (in short best hyperparameters).

Think of the map as finding many local clusters in the Iris-Versicolor class. Since SOM's behave somewhat like k-means clustering, I would say this implies Iris Versicolor has many local clusters which are as closely knit together as the entire classes of Iris setosa and Iris virginica.

Is it normal that SOM clusters the instances with the "versicolor" class into multiple different BMUs?

1 Answers1