Croissant RDF roadmap

[david4096]

There are a number of open issues. If you just want to discuss general direction feel free to add your thoughts or haven't quite figured out how to express it as an issue in this thread, anything goes! Some ideas for best next steps:

• Regenerate croissant ML JSON-LD from RDF
• Merge RDF from multiple providers

Bigger ideas that don't have an issue yet:

• Connect better with downstream applications (via @stefanches7)
• Annotating the metadata using other ontologies
• Querying alongside other knowledge graphs
• Getting a persistent Croissant knowledge graph online

Things that don't quite fit in this repo (but might?)

• Getting bioinformatics datasets represented in Croissant ML
• Getting ML tools easily using bioinformatics data (Rea demonstrated this)
• Connecting SPARQL to dataset acquisition > tensorflow tfds
• Creating composite datasets using Croissant metadata from multiple datasets

For annotating with EDAM, or connecting other ontologies I was given these suggestions

• use dataset description to map as well
• use a more broad set of ontologies
• combine CSVs from NCBO bioportal
• too many medical ontologies, how to choose, size concerns, chemical compounds
• concrete vs abstract classes, aim for the middle in terms of details. use abstract terms preferentially

Moved from david4096/croissant-rdf#46

[GeorgeDeac]

I think you guys might also need some sort of a rules and constraints intermediary ontology, that would aim to guide the merging process when combining multiple datasets.

[stefanches7]

@GeorgeDeac interesting point, I am currently thinking about what we could do for that. Maybe just take DCAT https://www.w3.org/TR/vocab-dcat-3/? Would be grateful for your input and recommendations!

cc @nolden

[stefanches7]

This one looks good too: https://rub-informatik-im-bauwesen.github.io/ir-ontologies/isoprops/isoprops.html

[stefanches7]

@david4096 what did you mean with "downstream applications"? Actual ML model runs?

I imagine Bioinformatics datasets going a quite similar way to what we do with OMERO david4096/croissant-rdf#49, except Bioinformatics does not have a unified data model (yet). GHGA Metadata Model may help us here though.
Could you give a link to where Rea demonstrated getting ML tools? Sounds super cool 👍

I really like the idea of ontologizing dataset description, it is of course always a guess if we'd do it ourselves instead of those who created the data. I would see it a good part of croissant editor, but worst case we could take it over too.

[GeorgeDeac]

Well you probably need something like https://www.w3.org/TR/shacl/ on top of it, to define ontology "shape" constraints. I was thinking this would be beneficial especially if there are highly nested datasets that might have constraints within groups. I previously had recently a rodeo with constraint ontologies on top of scene description ontologies :)

[stefanches7]

Thanks @GeorgeDeac , indeed just discussed SHACL in the Croissant WG meeting yesterday: https://docs.google.com/document/d/1OINP9AmphhAa3J9sw87QtpY9Hb4_Ym9Zxf1_ycF1_04/edit?tab=t.0#heading=h.mv4d0rbovaa0

Your rodeo sounds interesting, do you have the code of what appeared out of it somewhere open? It would be beneficial for us to consider corner cases, to design a more generalist system :)

[GeorgeDeac]

Nah not yet, it's still ongoing early stages, I've just passed the specification stage and would still need to consider some things.
You guys should just run inference on pyshacl after you think of a "macro set" of shacl specifications and then need to think how to handle the possible multiple choices, or how to prompt the user. Would like to contribute nontheless, seems interesting.

The really interesting thing I see with this in the future would be the ability to mine the metadata of a dataset with some text to semantic graph models.

[stefanches7]

@GeorgeDeac Idk if https://github.com/mlcommons/croissant/tree/main/editor already does something of what you mention here? 🤔

Yes, such mining is really a cool idea, but idk if it is not too generalist for the current roadmap - here we just want to pool the freaking triples from CroissantML JSON-LD providers, maybe annotating stuff with ontologies a bit.
But what you say could be achieved with LLMs and schemas in my opinion, I am personally really interested in that too.

[GeorgeDeac]

Yes and no, the mining part would be quite hard and low recall for most transformers, especially if the information is spread out (I've dealt with the same problem in my project where I attempted to mine specifications and turn them into ontologies).
A contrastive text to graph model, something like TAGA would be more appropriate, especially if it is finetuned on a specific ontology on the graph decoder part, so it knows what to expect and look for.

And yes it does to some degree, although it would need to be adapted for the shacl part.

[stefanches7]

Thanks for the TAGA pointer! Yes, we could try this approach too, once we have exact data. Here david4096/croissant-rdf#49 I have worked on adding microscopy data to the Croissant, and later to our graph. It has an underlying OME Schema, and we could use some bio ontologies (e.g. EDAM-BIOIMAGING).

[GeorgeDeac]

Yeah for sure, I would like to help on that one

[david4096]

Bioschemas FAIR Checker for Croissant metadata via Leyla

https://fair-checker.france-bioinformatique.fr/

[stefanches7]

Cool thing - Science DB (Central Chinese scientific repository) uses Croissant!
Example neuroscientific study:
https://www.scidb.cn/en/detail?dataSetId=f512d085f3d3452a9b14689e9997ca94#

I suggest we add them to the graph

[stefanches7]

I think you guys might also need some sort of a rules and constraints intermediary ontology, that would aim to guide the merging process when combining multiple datasets.

https://docs.google.com/presentation/d/1-9sg1X8siHZCa4Zh_7vIOzreEz0-3d-d/edit#slide=id.p35

DDI-CDI might be another standard to look for in this regard. Note how they want to build a Knowledge Graph themselves on the slide I have linked - maybe we could make sure our KG and their are connected in a meaningful way @david4096

[david4096]

We are discussing using this tool for creating a "dataset of datasets" as part of the MLCommons Datasets WG. Received this comment regarding that in the latest Datasets WG call

Steven De Costa
12:18 PM
That dataset of datasets could be a CKAN instance perhaps, see: https://www.ckan.org/blog/bridging-ckan-and-machine-learning-introducing-support-for-the-croissant-standard