som_quality_measures

A collection of algorithms for measuring the quality of self-organizing maps. This repository contains the following implementations:

1. Dx-Dy representation is a relatively simple algorithm for measuring the quality of a self-organizing map (SOM). The user can find all the details about the Dx-Dy representation in [1].

Dependencies

• Numpy
• Matplotlib
• Scipy

Platforms where the code has been tested

• Ubuntu 20.04.5 LTS
  • GCC 9.4.0
  • Python 3.8.10
  • x86_64

Example of use

If you'd like to estimate the Dx-Dy representation on your data, you will need to store your feed-forward weights in a Numpy file and then use a command like the one below (do not forget to replace the file name and the rest of the parameters):

$ python3 src/som_dxdy.py --grid-size-x 16 --weights-dim 2 --file ./data/weights_sample_noise.npy

There are two files, weights_sample_noise.npy and weights_sample_perfect/npy in the directory data/. You can try out the script by calling it upon those data files. For the first file, you will see that the Dx-Dy representation is spreading all over the plane; for the second file, it is aligned over the line x = y. This means that the first file contains a SOM that's not organized correctly. On the other hand, the second SOM has perfectly organized and captured the input space.

References:

1. P. Demartines, "Organization measures and representations of the Kohonen maps", First IFIP Working Group 10.6 Workshop, 1992.