The use of formal measures for the training of hierarchical Kohonen maps

We present a means of evaluating the relative recognition rate of unsupervised classifiers. Self organizing maps (so called Kohonen Maps) have some important features. They can classify data, reduce the dimension of input vectors and can be trained using unlabeled data. In our efforts to develop an automatic, context dependent classifier we highly relied on these features. Using unlabeled time series data, we had no way to compare different topologies with respect to “correct classifications”. We solved this problem by introducing a formal measure we called pseudo classes. This is an elegant, but heuristic method which can also be applied on other fields with context-dependent data