Abstract:

We compare space quantization obtained by codebook vectors yielded by Kohonen’s self-organizing maps and the neural gas methods. The average approximation error (called also the quantization error) is smaller when applying the neural gas method. The results are visualized graphically by scatterplot matrices. For the special case of 3-dimensional data the results, i.e. location of the representative codebook vectors among the original data points – are also visualized by a rotatable 3-D plot. The analysis was carried out for the Kefallinia data counting n D 3420 data vectors, each with p D 3 components. We stated that the results depend on the way of standardization of the data.