Figure 3

Comparison of K-means and Neuralgas using the Rand measure. This figure visualizes the problems of K-means in terms of reproducibility. Both K-means and Neuralgas were executed 25 times with a fixed cluster number of 20. In each repetition, the initial cluster centers were randomly sampled from the input dataset. A pairwise degree of similarity between each two clustering results was then computed using the Rand index. A value of 0.0 indicates no similarity, while a value of 1.0 means identity. The shown heatmaps visualize this measure. The upper part of this figure displays the results of K-means, the lower part those of Neuralgas. In all cases the outcomes of two invocations of both algorithms are slightly dissimilar, and not identical. However, the K-means approach reveals a significant higher variance in its results.