Data Mining

[riannella]

Is "data mining" the best term for the top-level concept?

[xfq]

Data mining is about discovering patterns, trends, and correlations in data, more focusing on analysis and discovery.

AI training is optimizing model parameters using data, focusing on building the model.

AI augmentation (is it inference? or fine tuning? does it include embedding chunks in RAG?) is "using asset(s) in the context of input into trained AI algorithms and models", which isn't data analysis, but rather data being consumed as input.

It seems somewhat inappropriate to categorize the latter two under the former.

[riannella]

Yes, I agree. What about the below? It has broader concepts (actions) and would enable more to be added over time?

[xfq]

I think it's better, but I think "inference" is clearer than "augmentation".

[riannella]

I am not convinced that "inference" is the right concept, as that is steps in logical reasoning, moving from premises to logical consequences. (It maybe that the AI Community wants to usurp this word :-)

"RAG" is the most common term for this, as in:
https://en.wikipedia.org/wiki/Retrieval-augmented_generation
Hence the "augmentation" action.

[xfq]

Yeah, while "augmentation" fits RAG perfectly, it fits poorly for some other model tasks. For example, if a model classifies an image as a "cat" or predicts a house price, is it "augmenting" the image/data, or is it "predicting"?

What about terms like "prediction", "execution", "generation", or "forward-pass" (this one may be too obscure)?

[riannella]

My understanding is that "augmenting" is when you use your Asset with an existing trained model.
(somehow it uses your Asset to generate "better" outcomes!)

So the "Augmenting" action is not about the final outcome, but about the action of using your Asset for this purpose ....

[riannella]

Updated Proposal ... comments welcome !!!

[aisaac]

Hi, sorry for the completely outsider feedback... But somehow I'm still receiving notifications, and interested by it.

Anyway, this discussion to me illustrates that it is odd that ODRL would become an authority to describe the various fields of (AI) application. Isn't there some better placed authority? Or isn't it possible to just let some consensus emerge within communities (e.g. some Data Space Support Center effort)? It seems likely that the discussions you're having now would hinder the adoption of whatever vocabulary you come with, because some other experts are going to be convinced that there is some flaw they don't like in it.

[riannella]

ODRL does not claim to be an authority on "AI" (I assume your comments also apply to IETF/CAWG/PLUS/RSL..?)
However we are really good at usage policies...(I hope that is an authority you agree to...)

If there is an SME group you are aware of, and they have a "taxonomy" of AI concepts, then we would like to know them.

We did look around and found nothing. The ones we reference also did not indicate the source of their proposals.

As we all know about standards, there is always an expert somewhere who will not agree with you :-)

[aisaac]

Hi @riannella , first I'm sorry for hijacking this issue for a more general topic than the starting point. Please tell me if there's a better space...
Then I think I would question the authority on usage policies. I mean, not about everything usage policy-related, of course ;-) but at least about identifying all applications some policy could consider. As you rightly point out in https://w3c.github.io/odrl/ai-vocab/, many have come with proposals, why would we have one more, which has the risk of the well-known https://xkcd.com/927/ ? Why actually not having chosen the one put forward by Open Future?
Or to consider it from a different angle: I've been involved in a project that used a category of AI tools/processes, which can be seen in the "AI categories" at https://ai4culture.eu/resources?page=0&resourceType=all . Why wouldn't ODRL embark in providing a vocabulary for these more specialised categories?

You might say that you inherit the scope set by the initiatives you've looked at, and I would very much agree. But then I'd say that you could inherit their semantics too. I believe that there is value in trying to 'align' existing proposals with the ODRL framework, as as been done earlier (I remember having seen something like this for CC licenses, but I can't find it now). Something like a very thin layer (a 'copy' in a way) of constructs that would allow you to say, 'if you want to use the IETF approach, then use this value, it's meant for it'.
But right now the wording of the ODRL profile ('has similar semantics') plus the diagram hints that it somehow tries to redefine all these terms. While you could try to term a URI for 'data mining' that's just equivalent to one of the approaches, and if another has a different/incompatible view on data mining, then it would get its own URI too.
Basically it would be rather to have the ODRL profile as a 'meta' view of these competing vocabularies, rather than being a competitor itself.

In that approach, one wouldn't need to define 'data processing' as top concept in the vocabulary. If it's a vocabulary for opting out of certain AI applications/processes, then the top concept could be as semantically neutral as 'AI-related processing' or something in the same vein.

[riannella]

We did look at the Open Future proposal: https://openfuture.eu/wp-content/uploads/2025/03/250307_Vocabulary_for_opting_out_of_AI_training_and_other_forms_of_TDM.pdf
(see page 10)

Our observation is that "Data Mining" is not a superclass of "AI Training".
(PLUS also does this)

We wanted the ODRL "AI model" to follow a model that represents the industry practice and can be expanded along those lines.
(An principle of ODRL is to document an information model of the domain.)

In addition, and in many cases, we can't reuse "exisiting proposals" as they do not publish URIs.

When/If they do...then we can added semantic OWL/SKOS relationships.

On Page 2 of the Open Future document, there are 13 proposals listed. They are all competitors.
Now there are 14. So the xkcd cartoon was almost spot on :-)