K-means and other pure clustering algorithms try to find the centers which best reflect the input's structure. Self-organizing maps have a different priority; the centers it is fitting are connected together as part of a two-dimensional grid, and influence each other as they move. The result is like fitting an elastic 2D grid of points to the input. This constraint results in less faithful clustering -- it is not even primarily a clustering. But it does result in a project of points onto a 2D surface that keeps similar things near to each other -- a sort of randomized ad-hoc 2D map of the space.
@author Sean Owen @since 1.0 | |