The quality of the map display should generally be evaluated by an expert in the application area. If samples having known classes are available, it is potentially useful to try to classify the samples using the map. Each map unit is labeled according to a majority voting of the samples, after which all samples that are projected into a unit are classified according to its label. The classification accuracy then indicates how well the classes are separated on the map, and the classification accuracy of new samples measures the generalizability of the result. Classification accuracy was used in this manner in Publications 1, 4, and 6.
The generalizability of the results could also give some indication of the quality of the mapping. Generalizability could be measured as the sensitivity of the map to small variations in the input data, caused either by adding artificial noise or by cross-validation. For the purposes of exploratory data analysis the sensitivity of the visualization should be measured instead of the sensitivity of the reference vectors of the SOM as such. In Publication 7 one possible sensitivity measure is presented; it reflects the difference in how two maps represent visually the relations between pairs of data items.