All of the analyses were performed using training data sets formed from the actual 29 opinions. First without and then with fuzzification. Also, the significance of either inclusion or dropping of the "No firm opinion" variables and the impact of preprocessing on training data were investigated. First by applying equalization and normalization to the complete data vector and then just normalization to each of the three variables formed from each statement. The original values, 1 to 4, could have been converted into a binary vector, too.
Figure 8 b) shows six different fuzzy variables of which the ones on the left and the ones in the middle seem to behave in a similar way in certain parts of the surface. For example in the left ones, there is a region at the furthest away corner of the surfaces in which the class membership values are large. Variable 
is an example of a variable, which is large in most neurons. This indicates, that most people have disagreed with statement 11.
Figure 8: a) The principle of 3D surface visualizations of SOM. b) A few variables visualized in 3D (the x and y are the SOM axes). Please note the different viewing angle of variables 
and 
.
Other similar areas can be found by checking the distributions of the values of other variables. This can be performed by selecting a few neurons to a new group and then visualizing the values of other variables as a bar graph, for example. Or the neurons can be grouped automatically by using simple rules. The easily located groups for our last analysis are shown in figure 9.
The last analysis included only 58 variables belonging to the "I agree" and "I disagree" classes and normalization was applied to each group of three variables for each separate statement. SOM training was performed with a neural network training algorithm, in which the data vectors are assigned to the best matching unit without trying to make the hit count of neurons match the distribution of the actual data. The distribution matching rule was used in all other analyses.
The resulting groups were formed using rules in which a class membership average of 0.7 was required. After the groups were formed, all the available weight variables were visually verified against them. The meanings of the groups were not evaluated by a professional of psychology. They are just examples.
Figure 9: Formed groups in the last analysis case. This figure illustrates a 2D 16 by 16 neuron SOM.
As an example of the conclusions that could be drawn we could try to understand the lower right corner group labelled with Oa. There are several negative opinions dominating the answers. A negative opinion to 02, 07 and 09 indicates, that these children do not appreciate the work done by their teachers. 03 indicates, that they do not enjoy being at school, and 24 additionally points out, that they dislike homework. This could be the profile of a problem child.
The formed groups seemed similar in the other analyses. Even the hit counts almost matched in each analysis.