The DataCrate specification for packaging research data
About this project
This project is to develop a specification for a standardised way of packaging file-based research data for re-use and distribution. (Archiving and preservation are not immediate concerns but we want to work with those communities of practice ASAP).
The current draft version is 1.0
Peter Sefton and Peter Bugeia developed a way of bundling data from the [HIEv] repository for deposit in the university Research Data Repository and for re-use by researchers. This work was presented at eResearch Australasia in 2013. The idea was picked up on the Cr8it project with which we were both involved, but without the benefit of a specification to align the two pieces of work. Two other projects, [Collections] and Calcyte have also embraced the idea.
It's time for a specification which we can all agree on for the structure of Data Crates, and which metadata standards to use and get on with creating tools that help researchers.
In this context, packaging data means:
Being able to distribute a data set via a single file (using Zip, TAR or a disc image format as appropriate) and/or via a URL with integrity checks
Host a data set on a web server with appropriate access controls with:
- An HTML page containing summary metadata down to (at least) the file level
- Machine readable linked-date
The goals are:
- To maximise the utility of the data for researchers (including the original researchers' 'future selves'). Given that a researcher has found a DataCrate package they should be able to tell what it is, how the data may be used and what all the files contain.
To enable discovery of the data by exposing metadata as widely as possible (respecting access rights)
To enable automated ingest into repositories or catalogues such as ReDBOX.
NOTE: The package format does not need to be able to be created by people, that will be done by software tools such as Cr8it and Calcyte.
Why are you inventing a new format?
We're not really. Data Crate is based on the Bagit packaging spec, with additional human and machine readable metadata in JSON-LD.
Why Schema.org? Surely there are suitable general purpose metadata formats for Research Data
Well, actually no there aren't any general-purpose, linked-data-ready, well-supported, extensible schemas that cover all of the kinds of general purpose metadata you might want to add to a research data package, and its component files, before sending it off into the world.
DCAT is close to what we wanted for this project but does not have coverage for all the fields we wanted at a high level, and does not handle the structure of datasets. DCAT also uses terms from a wide range of vocabularies, making it more complicated to use than Schema.org. It is RDF based, though. There is an effort to extend DCAT but it will take some time to complete.
Research Object Bundles (RO Bundles) are very close in spirit to Data Crates, and there is even an attempt to map them on to BagIt. However, RO metadata is very focussed on highly nuanced distinctions between different kinds of creators of research data using the PAV, without addressing a lot of general who-what-where type metadata.
The PROV ontology has been used to describe research activities, but its data model is too complicated for simple data modelling. It does not allow us to express "some entity created this file". In Prov activities create files. It is not practical to model every authoring relationship with an intermediate act of creation.
Datacite has a lot of metadata that is useful for Data Crates, but it is oriented towards describing things at the dataset level, not the file level. It is also focussed on published data which has a DOI (although you can enter a null DOI). There was an RDF version of Datacite but it is out of date.
RIF-CS is a standard created by the Australian National Data Service based on ISO 2146 for data dissemination which has good general purpose coverage but it's an XML based format that can't be used in a linked-data context directly.
Frictionless data packages are attractive in that they are simple, JSON based metadata but they don't interoperate with the semantic web, as there is no JSON-LD version, and they do not reference mainstream approaches to metadata such as Dublin Core except in passing. Having said that, cross walking between DataCrate and Frictionless data would be easy.
What about the Contents of files?
The initial versions of DataCrate do not concern themselves with the contents of files, beyond using Pronom file format descriptors. Future versions of DataCrate may follow the lead of the Frictionless data project, which allows for column-headers and variables in data files to be defined.
JSON-LD really? Isn't that a bit out there?
Remember, Schema.org, which is a standard sponsored by major web companies has as JSON-LD as the recommended encoding. The commercial web has been using this approach for years.