1

As we know from K-Means after the sample data are clustered in N clusters (each cluster has a centroid vector) not all the data are clustered in the clusters that they belong to! I mean that some data vectors may be clustered in wrong clusters. This means that even in K-Means there is not a 100% precision while clustering. I was wondering if such an "error" occurs also in SOM algorithm. So...after the SOM algorithm converges are there any data samples that do not belong to the node that they are actually put?

I hope my question was clear enough. I look forward to your answer.

Drill
  • 169
  • 2
  • 3
  • 12

1 Answers1

0

SOM clusters data according to the inputs presented. The formation of the clusters is dependent on the way the inputs are presented to the network.

The SOM clusters the inputs by minimizing the distance between a winning node in the map lattice and the actual input vector followed by an adaptation on weights in the immediate neighborhood. So the inputs that belong to a single class gets mapped around the same location forming a cluster.

So when the algorithm converges it is very unlikely that this would happen unless in the following scenarios

  1. Not enough data points were presented to train the network
  2. Value of learning rate used is not proper
  3. SOM clusters correctly and we do not see the common(hidden) features in the clustering and tend to think there is a mistake in clustering (happened to me many times).

I'm not an expert of the SOM but answering from experience. Hope this helps.

Synex
  • 207
  • 6
  • 18