Tech Notes

Q. What should I set as the neighborhood radius when clustering with self-organizing maps (SOMs)?

A. The neighborhood radius that you set should depend on the 2-dimensional layout of the SOM grid structure. The neighborhood radius determines how many nodes will move together towards a sampled gene, in grid distance (measured by Euclidian distance). For instance, if the neighborhood radius is set to 1, all the nodes within a grid distance of 1 will move together with the primary node.



The higher the neighborhood radius, the more nodes will move together. Based on the same idea, setting the neighborhood radius to 0 will result in completely unrelated clusters.

As a rule of thumb, set the neighborhood radius to no less that the smaller grid dimension and no more than the larger dimension. For instance, if you choose to perform a 3x4 SOM, you should set the neighborhood radius between 3 and 4.