People in Machine Learning have been asking themselves since ever how to use and how to built structures.
People in Knowledge Acquisition build "ontologies" that are structures. Data Analysis developped plenty of techniques for classifying data in structures, all based on a definition of the distance between two clusters. A recent incarnation of these techniques has been acknowledged by the KDD community under the name of Birch. When no distance is available, classification measures have been developped to cope with the case. If generalization then see AQ adapatations to clustering, if "utility" then see Cobweb, if Bayesian then see AutoClass. Is a structural representation a set of theorems? (of the form ForAll x, subclass (x) implies Class (x), as dog(x) implies mammal(x))? Is not a kind of heritance property within the structure necessary? If you answer no, then what is a structural representation, as opposed to a non structural one? In this discussion devoted only to ETS? (sorry, I do not know this formalism) Cheers Yves
