Linked Open Data (LOD) is the largest, collaborative, distributed, and publicly-accessible Knowledge Graph (KG) uniformly encoded in the Resource Description Framework (RDF) and formally represented according to the semantics of the Web Ontology Language (OWL). LOD provides researchers with a unique opportunity to study knowledge engineering as an empirical science: to observe existing modelling practices and possibly understanding how to improve knowledge engineering methodologies and knowledge representation formalisms. Following this perspective, several studies have analysed LOD to identify (mis-)use of OWL constructs or other modelling phenomena e.g. class or property usage, their alignment, the average depth of taxonomies. A question that remains open is whether there is a relation between observed modelling practices and knowledge domains (natural science, linguistics, etc.): do certain practices or phenomena change as the knowledge domain varies? Answering this question requires an assessment of the domains covered by LOD as well as a classification of its datasets. Existing approaches to classify LOD datasets provide partial and unaligned views, posing additional challenges. In this paper, we introduce a classification of knowledge domains, and a method for classifying LOD datasets and ontologies based on it. We classify a large portion of LOD and investigate whether a set of observed phenomena have a domain-specific character.