Define input/output format

[cmungall]

I will write proposals as individual comments in this ticket

[cmungall]

RDF

Pros:

• already a standard

Cons:

• multiple ways of handling reification
• even if we bless one, still awkward to handle

recommendation: The format MUST have a defined mapping to the RDF model, but there MAY be another serialization format

[cmungall]

GraphML

Pros:

• standard
• allows for property graphs

Cons:

• no standard way of defining which node or edge properties are accepted

[cmungall]

Translator TSV

Spec (loosely defined):

• TSV or CSV
• Nodes and Edges as separate files
• Multivalued columns separated by |
• Must be readily translated to a python DataFrame - i.e. first line should be a header
• Headers must come from Translator spec http://bit.ly/tr-kg-standard
• E.g.
  • node
  • edge

[cmungall]

Translator JSON

Follow KB standard

Structurally almost identical to TSV above, the doc would have a section for nodes and a section for edges

Cons:

• larger I/O or disk footprint than TSV for little gain

[RichardBruskiewich]

I've expressed this reservation before (to you Chris), but I'm wondering whether simply defining the nodes and edges alone suffices for knowledge graph representation. In effect, we need to annotate statements, not just subject nodes of a statement, but simply annotating the predicates doesn't help either. It does seem to be necessary to treat statements as a reified node, then hang everything off of it: subject, predicate, object, evidence, provenance, etc.

[yy20716]

Chris, I wonder if we could also briefly mention GraphQL and Tinkerpop as well. The data model used in these languages/systems are also based on or similar to the property graph, so their pros and cons are also very similar to the ones of GraphML. I listed down additional pros and cons as follows.

• One of the advantages of the models based on the property graphs and their variations is that they may be not suffered from the reification problem in RDF. The data can be presented and formatted in more compact ways (compared to RDF).
• However, employing the property graphs may be not a good choice for the case where we need to integrate or link multiple different datasets, which can happen in this project. For example, most property graph models do not use concepts of IRIs and typed literals for describing their entities, thus it can be challenging when we need to handle entities labeled with the same values, which often exist in different datasets.

[cmungall]

@yy20716 we will need to be careful about how we map CURIEs to IRIs. If the exchange format is not an RDF serialization (which already has precisely defined mechanisms) we will need to embed a prefix mapping in the exchange file or have a standard one.