Questions regarding to probabilistic RDFs

[adielstatman]

Hi.
I want to operate a real-world decision making robot with a knowledge graph which uses an RDF schema, and can deal with uncertainty and probabilistic data. However, regarding the uncertainty issue, I find the following concerns:

1. Seems that none of the main available knowledge graphs (Neo4j, AllegroGraph and GraphDB) explicitly supplies computations of probabilities. Can you help me understand why, and how can it be that such a crucial aspect is lack?

2. There were some former attempts for probabilistic RDFs, such as URDF (Meiser et al., 2011), or (Lian and Chen, 2011) but seems that none of them became useful in research and used as an operating or a supporting system. It seems that the common way for reasoning in knowledge graph is embedding, particularly the probabilistic methods such as UKGE (Chen et al. ,2019). Can you help me understand why?

3. Are you sure that your method is sufficient for this purpose? In this sense, what are the differences between this and the former attempts? Can it be integrated as an RDF reasoner in one of the mentioned knowledge graphs and add them the value of uncertainty and probability?

Regards,
Adiel.

[afs]

Are you sure that your method is sufficient for this purpose?

RDF-star isn't making that claim.

RDF can recorded information about probabilities but RDF itself is not a probabilistic computing framework.

RDF can say "X believes {:s foaf:name "foo"} with probability 0.9"
But the graph does not contain the fact :s foaf:name "foo".

RDF-star doesn't change this. It's in the nature of RDF.

RDF-star can help recording information -- e.g. triple "s p o" was seen in graph G. Even if the application A concludes "s1 owl:sameAs s" then RDF-star does not automatically imply "s1 p o" is in G. (After all, how does A communicate its proof that "s1 owl:sameAs s" if that proof involves "s p o" itself?)

What is true is that the information from G combined with A's knowledge/assumptions means that A believes G implies "s1 p o". Who ever created G did not make that statement and may or may not agree with the conclusion.

How some layer above RDF (or RDF-star) works with probabilities recorded in RDF is something else. As you note, there has been some academic work. RDF-star is a wider-community effort, not a research effort.

[adielstatman]

OK, thank you.
Can you recommend me on such "probabilistic engines" that compute probabilities?
Which are the methods that became useful in the industry or in real world implementations?

[puremachinery]

Pyro and TensorFlow Probability are two such frameworks used in industry. Good luck.