Analyzing the behavior of the SOM through wavelet decomposition of time series generated during its execution

Abstract

Cluster analysis applications of the SOM require it to be sensible to features, or groupings, of different sizes in the input data. On the other hand, the SOM's behavior while the organization process is taking place also exhibits regularities of different scales, such as periodic behaviors of different frequencies, or changes of different magnitudes in the weight vectors. A method based on the discrete wavelet transform is proposed for measuring the diversity of the scales of regularities, and this diversity is compared to the performance of the SOM. We argue that if this diversity of scales is high then the algorithm is more likely to detect differently sized features of data. © Springer-Verlag Berlin Heidelberg 2008.