Using n-grams for the automated clustering of structural models

Model comparison and clustering are important for dealing with many models in data analysis and exploration, e.g. in domain model recovery or model repository management. Particularly in structural models, information is captured not only in model elements (e.g. in names and types) but also in the structural context, i.e. the relation of one element to the others. Some approaches involve a large number of models ignoring the structural context of model elements; others handle very few (typically two) models applying sophisticated structural techniques. In this paper we address both aspects and extend our previous work on model clustering based on vector space model, with a technique for incorporating structural context in the form of n-grams. We compare the n-gram accuracy on two datasets of Ecore metamodels in AtlanMod Zoo: small random samples using up to trigrams and a larger one (∼100 models) up to bigrams.