Once a set of data fields is divided into categories, a Self-Organizing Map is trained for each category, giving a set of cluster prototypes. These resulting SOMs are then used to identify certain behavioral groups by collecting neurons that themselves hold a set of data vectors. Groups can be labelled automatically using prior knowledge and rules, or interactively by the user.
After the behavioral groups have been identified for all categories, each data record (answer form) is replaced with a vector of truth values, denoting the belonging to the observed groups. In most cases it is also sensible to have some kind of group membership for those data records that were not identified. These truth values represent answers from a particular point of view and are then combined with the concluding SOM.