When publishing Linked Data it is common to undertake the inter-linking of newly published data with existing sources as a final step in the publishing process. Linking strategies range from manual methods through to automated link discovery. The volume of out-bound links may vary over time, e.g. as new links are discovered to existing sources, or new links are made to other new datasets. Over time, due to "semantic drift", it may be necessary to remove links to some external resource.

The linking data within a dataset may therefore have different quality assurances associated with it, as well as a different rate of change to the core dataset.

Partition the data into two datasets: the core data and the linking data. Publish linking data as a separate set of documents. Use See Also links to tie the data together.

#Core Data
<http://www.example.org/places/Paris> a ex:Place;
  skos:prefLabel "Paris";
  rdfs:seeAlso <http://www.example.org/links/Paris>.

#Linking Data
<http://www.example.org/links/Paris>
  owl:sameAs <http://dbpedia.org/resource/Paris>.

  

Partitioning links from core data has a number of advantages in terms of both how it is published and how it is consumed.

For the publisher, partitioning links allows the linking data to be revised and improved separately to the publication of the core reference data. This is particularly useful where the linking data is generated automatically using heuristics. The quality of the generated links may need to be refined and iterated over time. As new datasets are published, the linking data can get updated to include additional relationships. Managing the typically smaller subset of links distinctly from the core data, e.g. in a separate triple store, allows more flexibility in managing the data.

For a consumer, the partitioning of links into a separate dataset allows more choice in what data is to be consumed. Because the quality characteristics of linking data may vary considerably from that of the core data, the ability to selectively consume only the core data is an important one. If the linking data is presented as part of the core dataset, then this requires the consumer to filter the data to remove the unneeded triples.