Appropriate neighborhood size for SOM

Question

I'm doing a program that can cluster numeric data using Kohonen Self-Organizing Maps and I'm trying to make it as generic as possible. So, how do I know the appropriate initial size of the neighborhood in proportion to the number of items (number of output nodes) in the dataset?

Suggestions would also be greatly appreciated. :)

Answer 1

You can refine the neighborhood function as you wish.

If you define a circle around each best matching unit (BMU) you can choose this radius to be proportional to the size of your grid in the low-dimensional space (likely 2d). Then you can make the radius remain constant or shrink according to a certain criterion until your last iteration.

Another option is to make the neighborhood fixed, as in a certain amount of surrounding nodes to the BMU.

Answer 2

if you want to have it "as generic as possible", there is probably only one possibility: random. But I'd suggest to make it "less generic as possible" and then you can play with parameters ...