Network Management Operations (nmop) WG Agenda - Interim on Knowledge graphs

• When: 2025-05-21 14:00-16:00 UTC
• Co-Chairs: Benoît Claise & Reshad Rahman

Detailed Agenda

1. Introduction (Chairs)

2. Education - Key concepts for building knowledge graphs

Raphaël Troncy presented slides.

Reshad Rahman> How would you recommend for the audience to go deeper in
this topic.
Raphaël Troncy> Links to EXCELLENT MOOC to learn Semantic Web
technologies:
https://www.fun-mooc.fr/en/courses/introduction-web-linked-data/,
https://open.hpi.de/courses/semanticweb
Raphaël Troncy> My second recommendation is to get your hands dirty

Boris Khasanov> Is Knowledge Graph a universal concept that can be used
for many things such as Networks, Vehicles etc.?
Raphaël Troncy> Yes, indeed it can be used for many things.
Boris Khasanov> Which tools can be used for creating the ontologies?
Raphaël Troncy> I.e. Protege
Boris Khasanov> Is RDF the most popular way to create/implement
ontologies?
Raphaël Troncy> Yes, RDF as graph model is often used for that and also
pushed by many big players for that purpose. We heard that the supply
chain of Amazon is managed via RDF.
Brad Peters> alternative to protege are Vocbench/showvoc
https://vocbench.uniroma2.it/doc/ which support RDFS and OWL
Ignacio Martinez-Casanueva> What do you think of the discussions of LLMs
v/s Knowledge Graphs? Are they complementary?
Raphaël Troncy> I believe they complement each other. For me merging
LLMs with KGs is the way to go. Plenty of ways to do it, active
reasearch area at the moment.

3. Education - Business justification of this architecture

Benoit Claise presented slides.
Brad Peters> That lack of semantic understanding across different
business units is also a potential security vulnerability as seen by
Salt Typhoon.

4. draft-mackey-nmop-kg-for-netops

Michael Mackey presented slides.

Brad Peters> From an OPS aspect, there is a lot of hetrogeneous data in
many different locations and hence the ontology gives and ability to
link the data together into a overall concept that aligns with Netops
needs
Ignacio Martinez-Casanueva> We are testing GraphDB database, which comes
with a free license
Brad Peters> Would be interested to hear from others on the high level
architecture and discuss the challenges and what IETF can contribute to
standardise the concepts and ontology potentially for re-usabiity
Benoit Claise> Brad: this is exactly what we want to cover in the last
section.
Raphaël Troncy> Yes, I would also mention GraphDB which has a strong and
scalable offering. You can add Neptune from AWS as well
Reshad Rahman> Why does LPG have no schema?
Raphaël Troncy> Because LPG is a data graph not a knowledge graph. They
went into representing data and later on they realized having a schema
would help. They are re-inventing the wheel.
Michael Mackey> For whatever reason, people tend to like LPG because it
is easier to learn if you are dealing with closed data

5. draft-marcas-nmop-kg-construct

Nacho Dominguez presented slides.

Brad Peters> LPG is easier to learn because it is data graph and there
is no abstraction....essentially a physical this relates to that...
Raphaël Troncy> To be complete on the Triple Store topic, MilleniumDB
may be worth to watch as well,
https://github.com/MillenniumDB/MillenniumDB ... open source project
that aims to be the Postgress of Graph database. It supports both LPG
and RDF graphs ... BUT, as of today, it is not yet SPARQL complete.
Boris Khasanov> If I understand correcty ontologies are not always
use-cases driven...
Thomas Graf> @Boris: Correct. It describes relationships. subject <
predicate > object. It is equally understandable for machines and
humans.
Michael Mackey> GraphDB is a good option as well but also only a single
instance ... almost all of these DBs are in memory DBs in a single
instance and scale to the size of memory available, you need to pay to
go beyond that with clustering.
Michael Mackey> They design with "Competency Questions" ... it's the
equivalent of UCS for the model. What questions do you want the model to
answer.

6. draft-tailhardat-nmop-incident-management-noria

Lionel Tailhardat presented slides.

Ignacio Martinez-Casanueva> @Boris, an ontology should be scoped to a
particular use case, otherwise there is a risk of modeling the universe
\:D
Ignacio Martinez-Casanueva> That's also the purpose of competency
questions, which aim to solve a specific problem.
Boris Khasanov> @Ignacio - yes, understand that. I mentioned that just
for the sake of clarity that there could be more universal ones than
just pure one use-case.

7. Brainstorming and next steps

Benoit Claise as chair> One question we are asking ourselves is whether
IETF (and NMOP) is the right place to work on this. Here is the charter
situation (slide 5 from chair slides). The charter is open for an
experiment like this but yes we would need to discuss with the AD.

Brad Peters> In short there is no concept that enables linking between
models.
Brad Peters> Create modular ontology models that can be reused plug and
play, items originally the paper for rtege comide plugin.
Thomas Graf> Sharing knowledge between whom for which reason.

Benoit Claise> Do we have somewhere on the Internet, a place like
https://www.schema.org/, where network related ontologies are already
available? Can we re-use something existing?

Raphaël Troncy> Probably there are pieces which could be re-used.
Looking at the 3 presentations, I see some tension in between. Nacho and
Michael are suggesting in their presentation to use already existing
YANG semantics to define ontology. Where Michael described how to define
one ontology and Nacho the pipeline needed. Where Lionel says no and
says that YANG is the meta part where YANG is being used to model
multiple ontologies.

Raphaël Troncy> This serves a discussion wherever we use existing
semantics in YANG to define ontology or use YANG to define new
ontologies.

Michael Mackey> That's not exactly what I was saying. This is just the
starting point. We have lots and lots of YANG models. The telecom
industry has been trying to build model driven engineering. There is
Open Modeling Framework (OMF), TMF API's, ETSI, different layers,
resource versus service and BSS layer. TMF is approaching the problem
space from the top where IETF comes from the bottom and there is a whole
propriety mess in the middle. There is conversion mapping all over the
place. What is missing are traces which describes the connections among.
Do we need a new model for trouble tickets if Remedy has already one?
Netcool already has one for faults. Re-use those models for taxonomoies
rather then ontologies. That's where I was coming from.

Benoit Claise> If NMOP wants to work on this and validate, what success
would look like in one or two years from now. Does it mean we have
knowledge, a mapping language, tools etc.

Lionel Tailhardat> Success is not to develop an ontology that fits all.
It is about capabilities. Sharing knowledge across practitioners to
improve network design for instance. Sharing knowledge on anomaly models
as another example. The goal is sharing knowledge where semantic models
have translation and reasoning capabilities. Everything else is cherry
on the top.

Benoit Claise> To be in the area directors shoes. Practically, what does
NMOP need to put in its charter to do that, and what are the
deliverables? Sharing is a noble goal but at the end what do we do? Is
the IETF the right place?

Lionel Tailhardat> One possibility could be that one describes anomalies
in their framework where another uses that shared knowlede and compares
the results wih their framework. Do you see the same behaviour and do
you see the same root cause. Do you have the same network automation
results.

Michael Mackey> I see two things here. I completely agree what Lionel
said about sharing knowledge. Being able to share knowledge is
important. But how do you do it. You need to agree on format right?
There is an existing discussion thread with Brad on it. Is it OWL? The
other point, it might affect multiple working groups, take the mapping
between L3NM and L3SM models for instance. That would be a very good use
case. Relationships between two models. Capturing the knowledge how the
two models relate. The other aspect is models outside of the IETF. How
to establish relationship to those models. Take Broadband Forum or TMF
as example. Is an RFC the right way? Probably not.

Brad Peters> In response to Michael's segway. Between RDF and OWL,
conclusions can't be reached (analogy is how to pick a programming
language). How to share knowledge and concept. It depends, some things
are easier in RDF where others are more easier in OWL.

Michael Mackey> How well it is implemented in open-source is another
factor to be considered.

Brad Peters> What is your usecase is also a factor. Things like
transitive properties.

Michael Mackey> You need to start somewhere. That is why on the email
thread I suggested RDF.

Thomas Graf> As a quick remark. I was thinking that for sharing
knowledge sharing between whom and for which reason is also important to
be considered. From there the usecases are coming from there we can
develop.

Benoit Claise as chair> Raising 2 show of hands. The first show of hand
is about the interest. I wonder on the first show of hands wherever it
goes beyond the authors of the presented documents. The second about
wherever IETF/NMOP is the right place. If not, has it to do because IETF
doesn't have the right processes yet.

POLL RESULT #1 : Is knowledge graph important to investigate?
-> yes (8), no (0), no opinion (1)

POLL RESULT #2 : s the IETF/NMOP the right place?
-> yes (6), no (4), no opinion (1)

Reshad Rahman as individual> I said no on the second because of the
missing processes.

Benoit Claise as chair> Should IETF adapt their process then?

Rob Wilton> IETF doesn't have the process yet and probably NMOP isn't
the right place. Is IETF or IRTF the right place? Can we scope the
experiment tight enough for NMOP? Or should we scope more toward
research for IRTF.

Ignacio Martinez-Casanueva> For the same reasons I said not today. IETF
first need to evolve. We have seen it with YANG data models. Maybe it is
not the right moment.

Benoit Claise as chair> To take all input into account. Thre are some
people who finds it interesting to explore knowledge graph capabilities.
The group of people might be constraint among the authors. There are
mixed feelings about IETF being the right place due to its rigidness on
producing RFC's. Reshad and I were wondering wherever the authors could
organize themselves in a design team.

Reshad Rahman as chair> I think we have seen from the first poll that it
appears to be important for the working group. Wherever IETF is the
right place or not. We will try to find a way how to move ahead.

Raphaël Troncy> I think we have here many people with the right
expertise and IETF provides an enviroment which allows colaboration
where existing IPR's can be declared and dealt with. I see NMOP as the
right place.

Michael Mackey> If IETF would not be the right place. Where else to go?

Benoit Claise> Valid point!

Benoit Claise as chair> We have 3 documents with experiments. Can the
authors to self organize together. Should it be schema.org or the IETF?
That would go a long way. I see people nodding.

Raphaël Troncy> IETF 123 in Madrid could be a good checkpoint.

Benoit Claise and Reshad Rahman as chairs> We will reach out on the NMOP
mailing list.