Latest / Agentic Conversations (formally mlops.community) / GraphBI: Expanding Analytics to All Data Through the Combination of GenAI, Graph, & Visual Analytics // Paco Nathan & Weidong Yang // #310

Transcript

GraphBI: Expanding Analytics to All Data Through the Combination of GenAI, Graph, & Visual Analytics // Paco Nathan & Weidong Yang // #310 · Agentic Conversations (formally mlops.community)

• 0:00
  So I'm Wei, my full name is Sacha Wei Dong Yang, but Wei is
• 0:04
  easier to pronounce And I'm Co for cannabis at data analytics,
• 0:10
  visual data analytics company. And I'd love coffee.
• 0:13
  I think the civilization starts with the invention of coffee.
• 0:18
  So I have to drink a coffee. I do add milk to coffee because
• 0:23
  the the black coffee is a little bit too strong for me.
• 0:26
  Welcome back to another Envelops community podcast.
• 0:29
  Today we are lucky enough to have not one but two graph
• 0:32
  experts who have been doing this for a very long time.
• 0:35
  I got schooled. I felt like I learned a ton
• 0:38
  about how to use graphs as tools and ways that we can leverage
• 0:44
  them better. Let's get into this conversation
• 0:47
  with Paco and way. As always, I'm your host
• 0:49
  Dimitrios, and you know what is a huge help?
• 0:53
  If you can hit a little review and whatever you are listening
• 0:58
  to this on that would mean the world to me.
• 1:00
  Boom. Let's jump into it.
• 1:01
  And Oh yeah, if you are one of those people that is listening
• 1:05
  on a podcast player, I have got the recommendation for you for
• 1:13
  our music reco. This is thanks to one of the
• 1:20
  people in the community, Lee Wells, who just joined.
• 1:23
  And now whenever someone joins the community, I asked them what
• 1:27
  their favorite music is. Today we're listening to We Are
• 1:31
  Won by Maze. We were talking about PII and
• 2:32
  using different methods to anonymize data, right?
• 2:37
  And Paco you, you had said something that I didn't fully
• 2:39
  understand and then way you said something else that I didn't
• 2:43
  fully understand. So maybe we can rehash that and
• 2:45
  I can understand it the second time.
• 2:49
  Awesome. Well, well, I, I was going to
• 2:51
  ask if you all ever came across, there's another podcast that I
• 2:54
  followed called the Dark money Files.
• 2:57
  And it's, you know, people who, there's a, there's a couple of
• 3:01
  consultants who have worked in banks and understand a lot of
• 3:05
  the insurance and outs of, of financial crimes and, and
• 3:08
  investigations. And so I was just going to
• 3:11
  preface it because they've had a great series recently.
• 3:14
  If you've ever heard of this thing called the SAR, it's a
• 3:16
  suspicious activity report. And the laws are really weird
• 3:20
  depending on what country the bank is in.
• 3:22
  But basically this, if you're at a bank and you see some
• 3:26
  suspicious activity, like there's a money transfer and the
• 3:29
  the counterparty is like a known terrorist group or something,
• 3:33
  you see something weird going on.
• 3:35
  OK, number one, you have an obligation to report a crime to
• 3:38
  a criminal investigation unit. If you see something suspicious
• 3:42
  and you don't report it, that's a crime.
• 3:45
  If you see something suspicious, you you have not an obligation
• 3:50
  but a responsibility to send it up the chain so that other
• 3:53
  financial houses might share. But if you send too much
• 3:58
  information, you might get sued. And then so there, there's these
• 4:02
  reports and it usually it costs on average about $50,000 to
• 4:06
  process each report. So you don't want to generate
• 4:09
  too many of them. And like machine learning models
• 4:11
  could generate thousands per day, which would be like, you
• 4:13
  know, 10s of millions of dollars of liability.
• 4:15
  So this whole space of like, what do I do?
• 4:18
  I'm getting, I'm getting attacked and what do I do?
• 4:21
  Because I mean, also these people are taking money and you
• 4:24
  might have to, they're under some situations as a bank, you
• 4:27
  might have to compensate if, if there is some kind of scam.
• 4:33
  So you could be losing money and facing like legal threats from
• 4:36
  three sides. And, and meanwhile there's this
• 4:40
  thing called a SAR and like, I've actually been yelled at for
• 4:43
  asking what I was supposed to integrate with something.
• 4:45
  I was like, can I see what the scheme is like?
• 4:47
  No, you're not allowed to. No, it's too confidential.
• 4:50
  So it's like it, it's just this whole tangle of worms about how
• 4:54
  to how to what do you actually do with with once you have
• 4:58
  financial evidence of financial crime or even suspicion of it,
• 5:03
  what next steps you take are really tangled.
• 5:06
  And I think wait on, you probably have a lot more
• 5:09
  experience about this in certain theaters too, so.
• 5:13
  I have some similar experiences where even the schema is not
• 5:17
  allowed to to see because it's the schema may actually view
• 5:21
  some secrets or certain activities may become liable to
• 5:26
  certain parties. So that's can be pretty tricky.
• 5:30
  And so it basically gives, it gives away information that if
• 5:34
  you were looking at it, you know, because you know the
• 5:38
  schema, you can guess a few other parts of this puzzle and
• 5:43
  get information that people don't want out there.
• 5:45
  The bank, the banks are using a lot of data that come from
• 5:48
  providers. There may be other cases where
• 5:52
  there's data that's coming from, say public sector agencies,
• 5:55
  crime investigations, there may be intelligence reports.
• 5:59
  And so there may be parts of the schema that are highly sensitive
• 6:03
  and only certain people are allowed to see.
• 6:06
  But you were saying that with graphs and anonymizing that PII,
• 6:12
  you're still able to gather insights, right?
• 6:15
  Yeah, that was cool. We were, we were just in a talk
• 6:18
  and Brad Corey from Nice Actimize was showing where like
• 6:23
  they're preparing to do rag and they were using I think Bedrock
• 6:26
  and they know that they've got a hot potato.
• 6:29
  They know they've got a lot of customer PII that just can't go
• 6:32
  outside the bank. So what they were doing is
• 6:35
  substituting PII with unique identifiers.
• 6:39
  They generate tokens, they generate on the fly and then
• 6:42
  they make the round trip after they've run through LLMS and
• 6:45
  made a summary and they they replace the tokens with the
• 6:50
  highly confidential material they just have internally.
• 6:53
  And so this is a way of being able to use some sort of
• 6:55
  external AI sources, AI resources, but still manage a
• 7:01
  lot of David privacy. That's cool.
• 7:05
  Yeah, I've seen it with, we had these folks on here from Tonic
• 7:09
  AI and they were talking about how they would use basically the
• 7:14
  same information but swapping it out.
• 7:18
  So if it is someone's name, they just changed the name so it went
• 7:22
  from Paco to John. And if it is a Social Security
• 7:26
  name number, they would swap out the Social Security number and
• 7:30
  totally randomize the number, but it still is a Social
• 7:34
  Security number. So you at the end of the day,
• 7:38
  you get almost like this double-blind.
• 7:41
  So even if you're a data scientist looking at the
• 7:44
  information, you can understand it, but you don't know if it is
• 7:49
  the true information that's going to reveal that PII.
• 7:54
  Interesting. Interesting.
• 7:56
  Yeah. Although I do see situations
• 7:59
  where even the structure, the document itself, it gets
• 8:05
  reviewed. It reviewed information that you
• 8:09
  do not want people to know like reviewed it, like in the
• 8:15
  investigation space, very often you do not want people being
• 8:18
  investigated, know that being investigated.
• 8:21
  But certainly information, even the structure is structured, the
• 8:24
  document being reviewed can become a problem.
• 8:28
  So so at some point I felt like the in house on Prem LM might
• 8:36
  might be a like necessary like especially just red news that
• 8:41
  like what is it that the M3 Ultra Studio with a 500 gig RAM
• 8:49
  Quran large dump models at 2020 token per second That could
• 8:55
  potentially, yeah, a interesting solution for that.
• 9:01
  Yeah, I mean for, for our end use cases, you know, like 60% of
• 9:05
  those are air gapped. And so you know, the, the
• 9:10
  largest chunk of that they're they're going to be a lot of
• 9:12
  like public, public sector agencies running in skiffs.
• 9:16
  So they they can't do any data out.
• 9:19
  Yeah. And, and, and there's there's
• 9:22
  good news for running really interesting LLMS on, you know,
• 9:28
  local hardware there. There's a lot of really good
• 9:31
  news. I will shout out to my friends
• 9:33
  over at useful Sensors, Pete Warden and and company.
• 9:37
  I'll put that in the chat. You can do a lot with with hard
• 9:40
  work with local hardware. Yeah, what?
• 9:42
  What are they doing? Useful sensors.
• 9:47
  So Pete Warden and Mentorath Kudler.
• 9:52
  They were part of the Tensorflow team at Google and for I think
• 9:57
  like 8 years, they evangelized use of deep learning inside of
• 10:01
  products at Google, like internally.
• 10:04
  And then they left and, and the team has a start up in Mountain
• 10:07
  View now. And what they're showing is,
• 10:10
  hey, here's like $50.00 worth of hardware.
• 10:12
  Here's, you know, here's an ARM chip with a neural network
• 10:15
  accelerator on it and we can run 3 LLMS on battery power.
• 10:20
  So it it, it's pretty cool because they they came out of
• 10:23
  like the tiny ML if you've ever seen the conference.
• 10:27
  Oh yeah. And so, you know, this is a lot
• 10:31
  of the specialty that Pete has and, and manager, you know, he
• 10:37
  was on the CUDA team at NVIDIA before.
• 10:43
  So I mean, these folks really know how to make AI
• 10:48
  infrastructure run on hardware and particularly how to handle a
• 10:51
  lot of low power and low latency kinds of situations and you
• 10:56
  know, where to punch through the bottlenecks.
• 10:59
  You don't necessarily have to have an A ginormous GPU cluster,
• 11:03
  although in some cases it helps, but especially when you're
• 11:06
  running inference, you can be running on much lower power and
• 11:08
  doing really interesting things out in the field.
• 11:11
  So wild. No, I know that we had
• 11:15
  originally wanted to chat a bit about this idea that I think way
• 11:21
  you had proposed and it's a little bit of a a
• 11:26
  differentiation on graph rag. And so maybe you can set the
• 11:33
  scene for us because I want to go deeper there.
• 11:37
  Yeah. I I I I run in dangerous danger
• 11:42
  of a pulling way for I fundamentally, I think with LM
• 11:51
  whole machine process information has changed and
• 11:55
  before LM everything is exact symbolic, like matching like all
• 12:00
  the APIs, you all the rigid data structures.
• 12:05
  Just think about blue, deep blue, deep blue and beat the
• 12:10
  chess. Everything is rigid knowledge as
• 12:14
  the rules and things. LOM changed everything because
• 12:18
  LOM start to understand things in the contextual base start
• 12:23
  understand fuzzy things and and it suffers the same weakness of
• 12:28
  the human being. Not exact like we we we glide
• 12:33
  over informations, we draw conclusions, we make lips make
• 12:37
  jumps, but on it seem on the on the at the same time, LMS
• 12:42
  ability to to to reason like human that that for me is
• 12:47
  fundamentally changed how we approach the computing.
• 12:51
  And so so in the applying LM to analyze document, my feeling is
• 13:02
  my analysis is now we can let LM work more like human rather than
• 13:06
  like machine. It we understand in the past
• 13:09
  that also implies what the data structure is preferred,
• 13:13
  preferred for LM, which might say, I would argue that a data
• 13:19
  structure, a data management that preserves as much contact
• 13:23
  information as possible, preserve as much nuance as
• 13:27
  possible and that the subtle nuances may come out to be
• 13:31
  important. And so, so I, I, I, I use the
• 13:36
  example of my wife is presenting the American tourist to Brazil
• 13:42
  gets invited to a horse party, says the party start at 6:00 PM.
• 13:47
  So, so as a good American guy, show up promptly on time at 6:00
• 13:50
  PM and the Hostess comes out still wrapped in the shower,
• 13:55
  shower tower and totally confused and so right.
• 14:01
  And it turned out over there when this is 6:00 PM is where
• 14:05
  the the Hostess start thinking about the party start like going
• 14:09
  on shopping, preparing food and getting ready.
• 14:12
  And the people usually don't show up until like two or three
• 14:15
  hours later. And so that culture difference,
• 14:19
  yeah, Pago, right. If we try to capture that in a
• 14:22
  knowledge graph and what kind of a what kind of construct allow
• 14:26
  us to capture those subtle cultural nuances there.
• 14:29
  And that might become important in understanding the document
• 14:33
  later. So I think that's the challenge.
• 14:35
  Yeah. Pago, you want to add something
• 14:38
  that I'd like to hear what you think.
• 14:41
  What? What from a perspective of
• 14:44
  natural language, something that the models bring in, but it's,
• 14:49
  it's kind of a nuance and I, I, I don't think it's talked about
• 14:51
  a lot. There's a very recursive nature
• 14:53
  to how we as people talk with each other and tell stories and
• 14:57
  share information. We, we do reference it in the
• 15:00
  sense of like going down the rabbit hole, Like if you follow
• 15:03
  a thread too far, you're kind of going down the rabbit hole.
• 15:06
  And there's this very recursive nature of how we think and
• 15:09
  especially how we express it certainly comes across in
• 15:12
  written language, although we tend to think of written
• 15:14
  language as something linear. There's paragraphs and sentences
• 15:17
  and then it can all be diagrammed.
• 15:19
  But when you look at the actual references that are inside of
• 15:22
  those sentences, they're making recursive calls throughout a, a,
• 15:27
  a story throughout somebody's speech or throughout a book.
• 15:33
  And, you know, we can try to linearize that and and come up
• 15:35
  with like an index or a bibliography, but at the end of
• 15:39
  the day, it's a graph and you get this very self referential
• 15:42
  thing in any text. And this is something that the
• 15:45
  LMS have really, I think, pulled out.
• 15:48
  And we were also just in a part of the talk we were in also Tom
• 15:52
  Smoker from why how was showing about how they they leveraged
• 15:56
  ontology, they leveraged schema and, and chase after information
• 16:01
  recursively. So I, I that that's just another
• 16:06
  kind of view on this, but I way I love how you all are
• 16:10
  approaching this. You have a very powerful view of
• 16:14
  of kind of relaxing the constraints upfront, but then
• 16:16
  having the context propagated through.
• 16:19
  I realized there's a important philosophical approach
• 16:23
  difference between East and West and the East of philosophy very
• 16:27
  much drive towards the nature of things.
• 16:30
  And it's important, which is that that's very curiosity about
• 16:34
  nature of things, that the desire to have a definitive
• 16:37
  definition of nature of something is led to the great
• 16:41
  scientific discovery over the past several 100 years.
• 16:46
  The Eastern philosophy very much on the all side is focused on
• 16:49
  the contextual focus on shifting, changing nature of
• 16:52
  things like the Chinese, the Bible, the Taoism, the Bible,
• 16:58
  Dao De Jing. The first verse it says stalk
• 17:01
  Adolfi Tangdao means if you name something you get it wrong or
• 17:06
  it's like it's not permanent. It's really focused on
• 17:09
  impermanence of things that focus on everything changes
• 17:12
  nature when contacts with other things.
• 17:15
  So so that is essentially a graph now the the you're putting
• 17:21
  both things together. So so OK, I have to say that
• 17:24
  that attitude towards like, oh, everything changed.
• 17:27
  Thus we cannot see anything that's everything is fuzzy is
• 17:32
  very much contribute to the the Chinese technology science
• 17:36
  developed very far in about 1000 years ago and stalled.
• 17:41
  And a lot of his attribute to this like philosophical like
• 17:45
  things like reduce a lot that a lot of curiosity in drive down
• 17:49
  deeper into the nature of things.
• 17:53
  However, in practical things, there's there's some practical
• 17:56
  application of that approach, which in today with LLM and the
• 18:00
  graph, we really see that it's like a great combination of you
• 18:05
  allow certain things to be drilled down, to be very
• 18:09
  definitive defined, to be clearly defined within the
• 18:12
  context and but allow a lot of information, contextual
• 18:16
  information stay fuzzy. So in fact, like Pablo, I felt
• 18:20
  like I'm really excited about like integrating sensing and our
• 18:26
  graphics are kind of as a solution together because the
• 18:29
  sensing helps to drive this definitive part.
• 18:33
  Once you have the definite report drew down named defined,
• 18:38
  it really speeds up to make a lot of assessment fast,
• 18:42
  definitive and precise as which was crucially important.
• 18:46
  But on the other hand, you allow this loose structure of
• 18:49
  information decomposed as a graph that you can easily like
• 18:55
  retrieve and without losing the nuances, the subtlety like in
• 19:00
  the in the in the culture, different things like you still
• 19:03
  preserve that. So those thing come together.
• 19:06
  I feeling is the one how you how you want to grant LLM to pro to
• 19:11
  create a precise accurate and know the limit know when it does
• 19:17
  not know not to make a judgment. I think that's also very, very
• 19:21
  important. So, so in my mind is they get
• 19:24
  the graph and AI right now is present opportunity to allow
• 19:28
  this Western way of drive to the nature of things and Eastern way
• 19:34
  of focus on the contextual information come together to
• 19:38
  work together to solve practical problems.
• 19:41
  So, so very well said. And, and you know, the, the
• 19:43
  challenge that we face is we don't really know what the
• 19:46
  downstream application will be like.
• 19:48
  We're doing investigation. We're doing some kind of
• 19:51
  discovery, whether you're trying to find, you know, money
• 19:55
  launderers or whether you're trying to find, you know, who's
• 19:58
  my best customer for this hotel. It's a discovery process.
• 20:02
  And by nature of discovery, you don't know what the answers are.
• 20:05
  In fact, in a complex system, you don't even know where or
• 20:08
  how, just you know, it's unknown unknowns, right?
• 20:11
  So by preserving that context, then you are sort of fortifying
• 20:16
  yourself. So that when the time presents,
• 20:19
  you'll be able to make the right discoveries.
• 20:22
  You won't have cut them off in advance.
• 20:24
  I think it's if you, if you, if you go back to like before
• 20:27
  relational databases came out, you know, you go back to some of
• 20:30
  the earlier writings from Ted Cod and one of his colleagues
• 20:34
  was William Kent who did a book called Data In Reality, if you
• 20:39
  go back to some of the early like 1970s thinking about data
• 20:43
  management, it's really interesting to see where the
• 20:45
  lines are drawn because in this Western view, so much of data
• 20:50
  management was about, let's have a, a data warehouse.
• 20:54
  Let's pretty much throw away the relationships.
• 20:56
  Let's focus on the facts. We have a lot of, as we were
• 20:59
  saying, a very Western view of like, I just want to know like
• 21:03
  millions of facts and I will piece them together with a
• 21:06
  query. I'm not, yeah, I'm not really
• 21:08
  interested in preserving the context.
• 21:10
  So I mean, I, I think we have a long history from like did a
• 21:13
  warehouse and going too far on the Western side.
• 21:18
  But what is interesting to me is the conversation that we had
• 21:22
  with Robert Caulk on here probably three months ago and
• 21:27
  how he said we've completely thrown out ontologies and for
• 21:32
  his specific use case that isn't the way that they wanted to go.
• 21:37
  And I wonder if you guys have thought through that and what
• 21:42
  that looks like, what the benefits are?
• 21:45
  And is it one of these things where you potentially are
• 21:48
  experimenting on those levels too?
• 21:51
  In my perspective, ontology is important, but you have to know
• 21:55
  the boundaries. Like I give a parallel into all
• 21:59
  the theory in the physics table, like Newton's law, like Newton's
• 22:03
  law is important. It captures important truth in
• 22:06
  the nature. However, like just like any
• 22:10
  physics, I'm a physicist, any physics theorist, The moment
• 22:13
  when the theory is as a proposed, it's very important
• 22:17
  fact, important concept is you're waiting to be
• 22:19
  disapproved. So you never accept as the truth
• 22:24
  of everything. You you you have a theory.
• 22:27
  What park is math scientist. So so I think he's he's also
• 22:30
  very familiar with the concept. When you propose a theory be
• 22:32
  test true, but you're always looking for situations looking
• 22:37
  for the boundaries where the theory was with stop to be true.
• 22:41
  So I don't think ontology is anything different.
• 22:44
  It's just like ontology needs to be very well grounded, need to
• 22:48
  be the the the context needs to be defined in within this
• 22:53
  context of this part. This ontology knowledge is is is
• 22:57
  is is real is is truth. The problem I see as a lot of
• 23:03
  traditional knowledge graph approach is people ignore the
• 23:06
  fact that ontology has to be confined within a specific
• 23:10
  domain. The moment you step out the
• 23:12
  domain, you have problem. And now.
• 23:16
  But the other thing is we think this domain ontology is
• 23:19
  fantastic. It help you to solve problems so
• 23:22
  much faster, so much precise. But again, just like as far as
• 23:27
  you, you can define the boundaries, define the domains.
• 23:30
  It's great. I, you know, I what, what Rob
• 23:36
  Calkin and Ellen Thornquest and others at, at Ask News, you
• 23:40
  know, what they're doing is they're, they're looking at new
• 23:43
  sources, especially regional news sources across the world.
• 23:47
  And, and they really are finding like hard evidence, you know,
• 23:51
  groundbreaking evidence on the ground.
• 23:54
  Literally, if you're doing ESG work and you're trying to do due
• 23:59
  diligence on a, on a company or a set of suppliers and you want
• 24:03
  to find out like, what are their operations really like over in
• 24:06
  that other country where they're based?
• 24:08
  And then you find out they're engaged in like, I don't know,
• 24:10
  child labor or something. And, you know, you, you want to
• 24:13
  make other arrangements before your shareholders find out.
• 24:16
  So I, I think with Ask News, you know, they're out and they're
• 24:20
  looking, they're working with those publishers and, and
• 24:22
  they're, they're collecting that news and representing it in a
• 24:25
  graph. And yeah, as you were saying, I
• 24:31
  mean you and ontology, ontologies really don't work
• 24:35
  across domains. You really want to focus more on
• 24:37
  like closed world within a domain, having a full enterprise
• 24:40
  wide ontology, nice idea, but I rarely see it work.
• 24:45
  And I think that in the case of like understanding news reports
• 24:49
  on the world, you don't know what the domain is in advance,
• 24:53
  You only know this is what is being published.
• 24:56
  And so I think by relaxing that constraint at Ask News, they're
• 25:01
  able to come up with a graph of like, here are things that are
• 25:03
  related. You can you can follow this
• 25:05
  evidence and you can find more historically about this area.
• 25:09
  I think those are very important.
• 25:11
  But ultimately it will be shaped by some kind of context, some
• 25:16
  type of shared definitions and ontology is, is really more
• 25:20
  about sharing definitions and making sure we're, you know,
• 25:23
  describing the same thing. Because I, I swear, you go to a
• 25:26
  big company, use the word customer in front of one VP.
• 25:29
  You know, in, in sales, it means something different to like the
• 25:32
  VP in charge of procurement. So even like the words
• 25:35
  themselves don't cross domains. The graph is basically our idea
• 25:41
  that we know that there's connections like if you, if you
• 25:44
  do have your, your operations data, but then you also have
• 25:48
  your, your like sales data, you know, there's some connections
• 25:50
  across there. It's not exactly the same, but
• 25:53
  some stuff is connecting. So graphs show where those
• 25:55
  connections are. But I think, you know, think
• 25:58
  about like the example of Google Maps, like there's different
• 26:01
  levels of detail. And of course, any video game of
• 26:03
  course, has this too. But you know, if you're taking
• 26:06
  satellite data and like trying to stitch together a map, you
• 26:09
  zoom in and you can see the beach and you zoom in and you
• 26:12
  see the car tracks and you zoom in further.
• 26:14
  At some point, you're going to get to pixels, right?
• 26:18
  And you zoom out and maybe you see this landscape of like a
• 26:22
  beach next to the ocean, but then probably you zoom out at
• 26:25
  some level and they've got like the name of the beach, right?
• 26:28
  So there's like a high level detail.
• 26:30
  I think grass are much the same. There are connections at the low
• 26:32
  level. Like Ask News is saying is like,
• 26:35
  you know, here's reporting from Zimbabwe.
• 26:38
  This is like the reporters on the ground.
• 26:41
  But then you zoom out and you're like, OK, well, you know, what
• 26:44
  impact does this have on our our supply network?
• 26:47
  Do we have to really make different plans?
• 26:49
  Is there going to be like a war breaking out that causes, you
• 26:52
  know, all those shipping containers to be delayed by
• 26:54
  three months? I, I think at some level you,
• 26:57
  you need to think of the graphs as sort of collecting higher and
• 27:01
  higher into more abstracted, more refined concepts, if you
• 27:05
  will. And so the stuff at the low
• 27:07
  level was kind of like, let's see how it all fits together.
• 27:10
  The stuff at a higher level, it's like, oh, actually we can
• 27:13
  maybe do some inference on this or we can use this to help
• 27:16
  structure other data that we're going to piece together.
• 27:21
  So Dimitrios, you you actually touch up a really big subject
• 27:25
  that things now in the exploratory process and it's
• 27:33
  come back with the questions knowing what question to ask.
• 27:36
  I'm offering is 8090% of the work.
• 27:39
  So a prescribed things to give you the answer often miss the
• 27:45
  point or miss the important subtlety.
• 27:48
  But the problem is that how do you discover the question you
• 27:53
  need to ask? And so in the way that our
• 27:57
  brain, our perception, our visual perception, our brain is
• 28:01
  a fantastic. I don't want to call the machine
• 28:04
  like or don't want to even call it tool, but has this great
• 28:08
  power of C patterns in, in, in, in information.
• 28:13
  Like we, we look at the out in the sky, we see the cloud.
• 28:17
  We have some concept. We still have some kind of with,
• 28:19
  with like your performer, I look at your performance, your dance,
• 28:23
  like their information being expressed without, without being
• 28:28
  able to verbalize, to define it. But you have to watch it to to
• 28:33
  feel that maybe you watch line up, you stop be able to describe
• 28:37
  it. You start be able to say, Oh,
• 28:38
  this is something's is there. So so in a way that what the
• 28:43
  graph does is the graph is a fantastic medium for
• 28:47
  visualization. You look at the information,
• 28:50
  express it just like hope I will bring like when we think about
• 28:54
  you, Demetrius, I've immediately think about Paco because we in
• 28:58
  the same pot in room together. So that's association.
• 29:02
  So this association of multiple pieces, information entities in
• 29:08
  the space, if you visualize in effectively, it help you to see
• 29:13
  the patterns, help you to see all the missing links, missing
• 29:17
  patterns, things that that get our attention.
• 29:21
  And then we start being able to formulate the question to, to,
• 29:24
  to, to formulate to, to the, to answer the question.
• 29:28
  So, so it, it like more than a tablet data structure.
• 29:33
  I have to see the graph really helps us to, to, to, to, to
• 29:38
  engage our brain in this way to, to spot important information.
• 29:43
  Just go like, like a go watch a dance performance.
• 29:45
  You you see something definitive happening, but you know it
• 29:51
  before you engage your like a language like logical thinking.
• 29:57
  Afterwards you start things that start concepts start to form and
• 30:01
  then you can start to build things around it.
• 30:03
  Oh. Dude.
• 30:06
  How cool is? That yeah, it's, you know it
• 30:08
  before you can express it in that way.
• 30:11
  Absolutely. Yeah.
• 30:12
  I think a lot of analytics workflow is work the other way
• 30:15
  around. We focus so much on building up
• 30:18
  the queries, build up the like programs to, to drive it to to
• 30:25
  to drive answer. But as Park and and we in the
• 30:31
  investigative space, we all know that too often getting the hint
• 30:37
  is 80% of the work. Like if if you know that you're
• 30:43
  being attacked, you know that the they came in through some
• 30:46
  vector, there's probably some set of machines that are
• 30:49
  compromised. You're not seeing that, you're
• 30:52
  seeing where you know, the bad things are happening, stuff is
• 30:55
  being stolen or whatever. So looking across your network,
• 30:58
  just building up a graph of like the associations of what's
• 31:01
  happening during an attack, there's some placeholders, there
• 31:04
  are definite questions that could be generated like which
• 31:07
  machine was compromised? Maybe I should fix that.
• 31:09
  So I think from the operational perspective, you know, I mean,
• 31:12
  you kind of have to think of, I mean, we do think about that,
• 31:14
  right? We we do think about like how do
• 31:16
  we identify those unknowns? But the problem is that the more
• 31:19
  complex the problem becomes, the more that those unknowns are not
• 31:23
  something that can really be charted.
• 31:25
  They have to be sort of poked at and explored.
• 31:28
  Yeah. And I think that's why weigh
• 31:30
  what you're saying with the graph being this visual medium
• 31:34
  that we can poke at and we can explore and it gives us a
• 31:39
  different perspective with which we can work with and wrestle
• 31:43
  with. The data is some something that
• 31:46
  I hadn't heard before, but it makes complete sense.
• 31:50
  That from a historical perspective in terms of data,
• 31:52
  you know, it's something to bring out would be to consider
• 31:54
  about spreadsheets because like spreadsheets are sort of my go
• 31:58
  to example of this is all in tabular form.
• 32:00
  It's very, very sort of, you know, left brain, everything is
• 32:05
  very buttoned down. But the thing about spreadsheets
• 32:07
  that you never see is there is a really complex graph behind it,
• 32:10
  and it only works because of that.
• 32:12
  But they never show that. They just show the tabular part.
• 32:15
  But all the real knowledge and dynamics and all the real
• 32:18
  information you're capturing in a spreadsheet is about those
• 32:20
  different dependencies in the and how that graph functions.
• 32:24
  Classic. Of course we don't see it
• 32:26
  because that would be absolute chaos for us, right?
• 32:29
  Mind blowing. The graph is this fantastic
• 32:32
  media for for this perceptive thinking.
• 32:35
  Well, the challenge is like when we talk about graph, I think
• 32:40
  that we need to really, really like the separate two things
• 32:45
  graph in the media of information capture and the
• 32:48
  graph in the media to help us to think they're two different
• 32:52
  things. Graph as information capture.
• 32:56
  The the sole purpose is to capture information as precise
• 33:01
  as possible, as complete as possible.
• 33:03
  You want to capture as much truth as possible.
• 33:07
  However, graph as a way of thinking, if you take the the
• 33:12
  raw graph captured preservable of truth.
• 33:16
  Well, the problem is we can only hold seven piece information.
• 33:20
  I will bring at any given moment will be overwhelmed by all this
• 33:25
  graphs like If we like think about it, I'll bring in our way
• 33:30
  like even the vector embedding. I called it a implicit graph
• 33:34
  because vector embedding gave you medium to compute the
• 33:37
  similarity. Effectively you can construct
• 33:41
  graph. Construct a graph.
• 33:43
  Exactly. If you can, you can manifest a
• 33:46
  graph out of it. And so you will see that the
• 33:50
  graph being captured at the layer at the stage that that's
• 33:54
  really designed to preserve the ground truth, as much truth as
• 33:58
  possible. But then you need a way to work
• 34:01
  the data into a form that we can easily digest with our
• 34:07
  perceptive power and that that that is challenge.
• 34:10
  This is also why I in my in my mind is a lot of graph.
• 34:15
  In theory people know the graph is how we think, thus is
• 34:20
  important. But in practice that is a
• 34:24
  barrier. And how do you reconcile the
• 34:26
  need between graph as information capture medium and
• 34:30
  the graph as ours to support our perceptive thinking medium?
• 34:34
  It's a very different thing. Just going back to what you were
• 34:41
  saying with, we can relate each other because we're on this
• 34:45
  podcast together. We've done stuff together.
• 34:47
  Maybe there's certain things that come up in our memories
• 34:53
  that are going to be the most pertinent to that graph that we
• 34:57
  have in our head, but it's never going to expand more than 7 hops
• 35:02
  or seven different parts of that graph.
• 35:07
  Have you ever, have you ever worked with there is like a kind
• 35:10
  of, I guess rubric might be a way to say it came out of
• 35:13
  Carnegie Mellon, out of CMU. Jeanette Wing had this idea of
• 35:19
  what's called computational thinking.
• 35:22
  And so it's sort of like a four step process of like breaking
• 35:24
  down a problem and then being abstracted back out.
• 35:28
  It's really powerful and I've used it a lot in courses
• 35:31
  teaching people. But I, I think that there, there
• 35:35
  may be something kind of emerging as like graph thinking.
• 35:39
  And so just to throw out like a straw man here, this is kind of
• 35:43
  thinking out loud. But one of the things that we
• 35:46
  see in like Fen crime in financial investigations is a, a
• 35:50
  kind of graph thinking, a four step process repeated over and
• 35:53
  over where, you know, you, you do your best to build out this
• 35:58
  graph and it might have hundreds of millions of nodes or billions
• 36:01
  of nodes or some ginormous number, something beyond human
• 36:04
  scale, beyond, beyond human comprehension.
• 36:07
  But then Step 2 partition. So like, can we break out this
• 36:12
  enormous graph into some areas of, of subgraphs of patterns
• 36:16
  that are interesting, like, hey, this, this looks like a really
• 36:19
  good customer or hey, this looks like a, a, a money mule, you
• 36:24
  know, fraud scheme. And so you go, you do this
• 36:29
  dimensional reduction then because you go from like 5
• 36:31
  billion nodes in a graph down to maybe 10 or 20 that are
• 36:34
  interesting. And so that's a, there are graph
• 36:37
  algorithms like Louvain or like, you know, weekly commit
• 36:40
  connecting components or there are different ways to get down
• 36:43
  to that scale. And in like in machine learning
• 36:46
  in general, we're looking a lot of dimensional reduction, right?
• 36:49
  So once you've got down to that scale, now you can use other
• 36:53
  graph algorithms, like maybe between the centrality or
• 36:56
  different forms of centrality to understand how are these parts
• 36:59
  connected. And gosh, maybe there's like one
• 37:02
  node in there who's orchestrating the whole crime
• 37:04
  ring, which is typically the case.
• 37:06
  There might be like a person with a bunch of shell companies,
• 37:08
  right? And they're doing fraud.
• 37:10
  So that's Step 3 is like leveraging certain types of, of
• 37:14
  graph algorithms to sort of think of page rank.
• 37:17
  Let's bubble up to the top, the parts that are probably first
• 37:21
  good steps to investigate. And then Step 4, put it through
• 37:25
  a, a, a work process. And I mean, if you're working
• 37:29
  with people in a bank, put it through case management tools,
• 37:32
  you know, a level a analyst gets assigned it, they go and they
• 37:35
  start poking around the graph, they do something interactive,
• 37:38
  they work with the visualization and they apply what they've
• 37:41
  learned. Or you may have some agents
• 37:44
  involved here too to help like summarize and and and Dick up
• 37:48
  part, but it's a workload. So it's kind of a a four step
• 37:51
  process of sort of graph thinking, if you will, that can
• 37:55
  be applied and can integrate people and also AI technology
• 37:59
  together. Yeah, I wonder add one more
• 38:03
  thing to Paco set. It's really, really important to
• 38:06
  be able to narrow it down to be able to identified things at to
• 38:10
  to reduce, reduce, reduce. But there's also another another
• 38:15
  aspect which is a simplification abstraction like very often when
• 38:20
  you capture the data, you don't really like the the domain or
• 38:23
  you don't need to know the future question.
• 38:25
  So, so the domain is wide but we look for the information answer
• 38:30
  the domain is narrowed. When domains narrowed for
• 38:33
  example like I call Paco as a mad scientist, at some point I
• 38:38
  can just refer Paco as a mad scientist.
• 38:41
  I don't need to add the information because mad
• 38:44
  scientist is Paco and that only about in the specific domain.
• 38:50
  So the reason I say that is because a lot of information
• 38:54
  when you when you domain wide it like I, I, I, I, I call it when
• 38:59
  you capture information. I prefer I call it a pure edge
• 39:03
  approach. Like in the graph, edge has no
• 39:07
  properties. It's just edge.
• 39:09
  It's just association. Anything you need, the property
• 39:12
  means the things you may need may need amended up on.
• 39:16
  Maybe you have something pointed to it or point all to it.
• 39:20
  You keep it as a node. Now, as you're thinking very
• 39:24
  often, like I know Paco, but I know Paco, this relationship, I
• 39:29
  can carry a lot of contacts in it already.
• 39:32
  I don't need additional information to to show to tell
• 39:36
  how I know Parker. It's just can be in there.
• 39:39
  I know Parker itself is sufficient.
• 39:42
  So what that means is that when we present like I know Parker,
• 39:48
  that relationship as a single relationship right in the data
• 39:52
  layer, there might be a tons, thousands or tenths of thousands
• 39:56
  piece information there, but they come out as A1 single piece
• 40:01
  of concise information. I think that is where where I
• 40:05
  think a analytical workflow or visual analytic workflow should
• 40:10
  be, is to be able to go from a very detailed, the broad, the
• 40:15
  big large information distill or aggregate down to a simple
• 40:21
  representation, but is grounded in that particular domain in the
• 40:27
  particular context. So for us to, so we can
• 40:30
  communicate, we can, we can communicate in simple language
• 40:34
  rather than carry a lot of information we had to I know
• 40:38
  Paco, that's it. We don't need to know how we
• 40:42
  know each other. Where do we know each other in
• 40:45
  certain contexts? Is it almost like the data
• 40:51
  underneath is is like an iceberg in a way and you knowing Paco is
• 40:57
  like the tip of the iceberg. You have that one piece of
• 41:01
  information, but then if you wanted to get more granular, you
• 41:04
  can go down and see the whole iceberg.
• 41:07
  Yes. Wait, could could we?
• 41:10
  Could we say? Them that you know, we, we pull
• 41:13
  everything, we can't get everything together.
• 41:14
  It's very noisy. We can go up different levels of
• 41:17
  abstraction, but to your point then we're going up levels of
• 41:20
  abstraction in particular domains like for purpose.
• 41:24
  So we have some shared definitions and and then we can
• 41:28
  start to say, OK, now let's let's do our levain partitioning
• 41:32
  or or whatever. We then we start to like drill
• 41:34
  down into subgraphs. It's like maybe a five step
• 41:37
  process. Yeah, even with the Lovain
• 41:40
  committed calculation or any centrality calculation, the
• 41:44
  graph has to be simple because very often I think the graph we
• 41:50
  talk about is I call it the multi graph, it's called it like
• 41:57
  multi domain graph, like it has different type of information in
• 42:01
  one graph. So computing a centrality in
• 42:06
  that kind of a hypergraph as a hypergraph, it's very
• 42:09
  challenging or, or what does it mean as a result if you mix
• 42:13
  human and emails and it's it's difficult.
• 42:19
  So, so, so that process itself to me is we already need to
• 42:23
  prepare our transform our grab data into a form that is
• 42:28
  suitable for that centrality computation.
• 42:33
  Yeah, very often. Like you have to already project
• 42:36
  into a specific domain for that competition to happen.
• 42:42
  Very good. That, that's what I was thinking
• 42:44
  is like the data that you have only becomes relevant once
• 42:50
  you've narrowed it down in a certain way and you're looking
• 42:53
  at a certain plane of that domain and you say, OK, now, now
• 42:58
  we're going to be focusing in on this plane.
• 43:01
  That's when certain nodes and certain data and certain
• 43:05
  connections become relevant because you're looking at that
• 43:11
  layer almost that in my head, if I visualize it, and we're
• 43:15
  talking about that Google Maps example again, you're diving
• 43:20
  deeper and deeper and you see different structures depending
• 43:24
  on the layer that you're looking at.
• 43:29
  And, and this fits very well with like data mesh kinds of
• 43:34
  concepts, you know, Shamak Dagani talking about how
• 43:38
  different domains share, you have to abstract, you have to
• 43:41
  come up with the relations. I think Chad also has the idea
• 43:44
  of like contracts, you know, where you have relations across
• 43:46
  domains. So you share some definitions.
• 43:49
  You have to, you have to condense down to that level
• 43:51
  before you can go across domain. So yeah, if if we if we use the
• 43:56
  domains in an organization to kind of guide when and where and
• 44:00
  how do we condense down, then we can really, really take
• 44:05
  advantage of this kind of abstraction.
• 44:07
  But it's, it's almost like I realized after I said it, the
• 44:13
  there's two vectors or there's 2 dimensions that you are looking
• 44:17
  at when you are zooming in or zooming out because you're
• 44:21
  playing on the field of granularity, but you're also
• 44:27
  playing on the field of the domain and what is relevant in
• 44:31
  that domain. So if we have that X&Y axis, you
• 44:35
  can get more granular inside of the domain, but then you can
• 44:39
  also just go on the X axis and change domains.
• 44:44
  And so that like a kaleidoscope, when you turn it you see a whole
• 44:48
  different set of relations. Yeah.
• 44:55
  And I mean, in an enterprise context, this gets really
• 44:58
  bizarre because, you know, you, you, the people in the domains
• 45:02
  that you depend on may not even know that you're out there.
• 45:05
  You know, you may be consuming from some log files from another
• 45:08
  application that are like totally driving your product.
• 45:10
  So like, can we have some sort of contract so that we know
• 45:13
  about each other? But yeah, scooting across the
• 45:16
  domains, that's the, that's the key challenge to like leveraging
• 45:20
  these kinds of, of technologies because usually you, you are in
• 45:25
  a particular domain when you're making the decisions.
• 45:27
  But for most applications, you have to combine a couple
• 45:30
  domains, right? So it's usually like there's
• 45:33
  something interesting going on between like sales and
• 45:35
  procurement or, or sales and marketing or, or, you know, some
• 45:41
  other business unit. So usually, often times you will
• 45:44
  have to combine. And do you then try and create a
• 45:50
  graph, two different graphs that are connected to each other or
• 45:54
  is it one larger graph? How do you look at it in that
• 45:57
  regard? Why federation sounds good.
• 46:01
  I think trying to have one ginormous graph is usually weird
• 46:04
  and and those projects usually don't ever end.
• 46:08
  But federating and being able to go across domains and say, OK,
• 46:12
  over there, let me let me send you something.
• 46:14
  I'd like to know what you can, what results can you bring back?
• 46:18
  So are you making a prompt in graph rag across a different
• 46:22
  domain? Are you making a query?
• 46:24
  Are you running some algorithm? Whatever.
• 46:25
  But there there's some kind of information transfer, but
• 46:28
  federation. Yeah, I, I I can talk about
• 46:33
  couple my my personal experience.
• 46:37
  First bring information to graph is a step forward step up
• 46:43
  because some information as a tabular format, it needs to be
• 46:47
  confined to a very specific definitions as pretty narrow
• 46:51
  domain graph is IEI. There's one example I look at
• 46:55
  the US flight record you can you can download from Department of
• 47:00
  Transportations. They release every two weeks
• 47:03
  after the the damn thing has 140 columns.
• 47:07
  I think 100 really, really wide. And the reason because it's
• 47:12
  because the flight maybe get diverted.
• 47:16
  Whenever the flight gets diverted, you you add about
• 47:19
  10-15 columns of the informations.
• 47:22
  So then you need to capture the the flight maybe divert more
• 47:26
  than once like it twice. Is it enough?
• 47:29
  No, some 3 * 3 is not No, some is a four times.
• 47:33
  So they actually have a 5 diversions.
• 47:36
  But if you have divert 6 times, too bad, it cannot exist.
• 47:41
  So, so that's the limits of of tabular format in the
• 47:45
  information capture. With the graph, you relax a lot.
• 47:50
  You, you naturally you can have 1000 diversions.
• 47:53
  I don't care. You can, you can just like the
• 47:55
  graph can, can, can keep amending to it.
• 47:58
  So, so that is really, really a big improvement with the graph
• 48:02
  to allow you to have a lot more flexibility in capturing the
• 48:08
  information. And the other thing is like very
• 48:11
  often in the tablet format is it's very easy, difficult to
• 48:15
  check the mismatches. Like we have example of that
• 48:19
  bringing two data set manager from two different, two or three
• 48:24
  different department in the same organizations.
• 48:27
  Everybody know other body, other other person's data has a
• 48:30
  problem, but you can't force other people to fix it.
• 48:34
  And but the with the graph, when you bring things together, you
• 48:38
  immediately see the mismatches and that.
• 48:41
  So we have one example of a company spend a couple years,
• 48:45
  they could not reconcile the data.
• 48:46
  But once they bring the data into the graph, they start to
• 48:49
  see the mismatch. In one month they they fix the
• 48:53
  data. Problem, but they start to see
• 48:56
  the mismatch because of the dependencies.
• 49:00
  Because once it now, because we let's see, you know, the records
• 49:04
  unique, right? But then when you link the other
• 49:08
  record together, you need to just see, oh, this record is
• 49:11
  actually duplicating other systems that they they recorded
• 49:15
  differently. Somebody made a mistake there.
• 49:17
  Yeah, we see that a lot for entity resolution where you
• 49:21
  think, like a Social Security number should be unique, but
• 49:24
  then you're bringing in data from some other sources.
• 49:27
  And there there was an application where maybe early on
• 49:31
  the product manager said, yeah, we need to collect the Social
• 49:33
  Security number. And then later on they said, oh,
• 49:35
  no, we can't do that. Just put it in, you know, a
• 49:37
  dummy number. And so now you've got like this
• 49:40
  data set that has, you know, 5000 instances of the same
• 49:43
  Social Security number. So once you start to put in a
• 49:46
  graph, you're like, wait, isn't that supposed to be unique?
• 49:49
  How come there's like this enormous node with, like all
• 49:52
  these things connected to it? Something's wrong.
• 49:57
  So it's really also a great way to figure out data quality
• 50:01
  issues, yeah. Although there's security, I
• 50:04
  mean, going back to what we were talking about before, if you are
• 50:07
  looking in financial investigations, if you're
• 50:10
  looking at sort of a criminal investigation, OK, maybe you've
• 50:12
  got some open data, like here's, you know, sanctioned shell
• 50:16
  companies or whatever. And then maybe you got some
• 50:20
  private information like customers.
• 50:22
  But maybe you've also got some feeds of like, Oh yeah, here's
• 50:25
  an active investigation. We're looking at these people.
• 50:29
  But then these particular people, they have, you know,
• 50:33
  immunity because they're diplomats.
• 50:36
  So like there's all these different levels of security and
• 50:39
  you, you start to pull it all together in a graph.
• 50:41
  You get a very comprehensive view.
• 50:44
  Maybe not everybody can even see that.
• 50:46
  Like, you don't, you know, you don't want the police officers
• 50:48
  who are doing parking tickets to know that, you know, XYZ
• 50:52
  diplomat might be investigated for a crime like that
• 50:55
  information should not go out. So where do you draw the line?
• 51:00
  Because the graph really brings it all together.
• 51:03
  But then how do you handle security issues?
• 51:07
  Yeah, the the the access control with the graph is automatically
• 51:11
  harder than the tabular the traditional database.
• 51:15
  Well, it feels like one of these, what you were talking
• 51:19
  about a way with the ways that you visualize it, you can almost
• 51:26
  create different access controls on the visualizations.
• 51:31
  So I, I don't know if you've thought through that in a way,
• 51:34
  but is that, is that kind of how you go about it?
• 51:37
  So fundamentally access control needs to be in the data
• 51:41
  management layer. Like if the database can can,
• 51:45
  can can support access control, you're great.
• 51:48
  We have. We actually however run into
• 51:50
  situation that database do not have the sufficient access
• 51:55
  control that supports business need.
• 51:58
  So in that situation we actually have to implement a filter layer
• 52:02
  in the in the in the data acts. When we put the data from the
• 52:06
  database depends on the roles and teams and we we actually
• 52:13
  prohibit some information from being accessed.
• 52:16
  But that's not a fundamental solution.
• 52:20
  Fundamental solution has to be in the data data management
• 52:23
  layer. It's a, it's a hard problem.
• 52:25
  I, I in previous work, which is more like knowledge graphs being
• 52:29
  used for large scale manufacturing.
• 52:33
  You know, one of the things we ran into is security access,
• 52:37
  because you, you take like procurement data plus some
• 52:41
  operations data, plus some sales data.
• 52:42
  Put it all into your graph. Suddenly you have a picture of
• 52:46
  of like how the company works, but it's like a really
• 52:50
  confidential picture. It's like maybe the board could
• 52:52
  see this, but nobody else in the company should see it.
• 52:55
  So there's a real power there, but there's always, always a
• 52:58
  risk. And how do you manage that?
• 53:00
  Is is a mind bogglingly difficult problem?
• 53:05
  I read a book talk about the certain like intelligence
• 53:12
  communities when they go to another countries.
• 53:15
  In the past, you use like falsified identities and but
• 53:22
  today is not good idea anymore because all the open source
• 53:27
  intelligence out there, even you want to withheld some
• 53:31
  information, but people can stitch together picture because
• 53:35
  of a related piece of information.
• 53:37
  Sit there outside on the social media, like maybe there's a
• 53:42
  picture of you with somebody that you did not take a picture,
• 53:45
  you did not post it, but somebody post on Instagram.
• 53:48
  And so all those information out there can can essentially is a
• 53:53
  graph can link back to you, even though you try really hard to
• 53:58
  stay, stay hidden at that. That's that's a fundamental
• 54:03
  thing problem in the in terms of the privacy security or you want
• 54:08
  to control the access information.
• 54:10
  But because you have all those connections the in in the graph
• 54:14
  that make it make it like really really hard.
• 54:17
  And, and, and a correlate with that, when I, when I talk with,
• 54:21
  you know, people in enterprise who are doing large scale
• 54:23
  knowledge graph practices, the one thing that I keep hearing
• 54:26
  over and over again is companies using graphs for market
• 54:31
  intelligence, or maybe sometimes you would say competitive
• 54:34
  intelligence, But you know, a lot of this might be for like
• 54:38
  sales win back strategies. I'm trying to understand who's
• 54:41
  the competitor that got our bid away from this?
• 54:44
  How can we go back and try to like, you know, give them a
• 54:47
  better quote? And so I've, I've heard this
• 54:50
  over and over again. We're like, that's one of the
• 54:52
  first graphs that starts making a lot of money is like literally
• 54:56
  doing intelligence inside the enterprise.
• 55:01
  Yeah, I was, I was going to go down that route of like, let's
• 55:05
  talk about a few other cool use cases that you have seen,
• 55:09
  whether it's just graphs or it is graph rag, which is a hot
• 55:15
  term these days, you know. I mean, I, you know, it's
• 55:21
  interesting, there's a lot of graph database vendors and they
• 55:25
  really kind of lean heavy on the on the graph query side of how
• 55:29
  to run this. And that's something's very
• 55:30
  familiar with people in data engineering, data science, you
• 55:34
  know, using a query. But I think in the graph space,
• 55:37
  there are other areas that aren't query first, like using
• 55:41
  graph algorithms or using, there's a whole other area of
• 55:46
  what should be called statistical relational learning.
• 55:49
  But you know, you probably heard of like Bayesian Nets or
• 55:51
  causality or, or different areas over there of using graphs.
• 55:55
  But then there's also graph neural networks, like how can we
• 55:58
  train deep learning models to like understand patterns and try
• 56:02
  to suggest, hey, I'm looking at like all the contracts you have
• 56:06
  with your vendors and I noticed that these 3 here are missing
• 56:10
  some terms. Do you, you know, is that a
• 56:12
  mistake? So I, I think that, you know,
• 56:16
  there's, there's the queries, there's the algorithms, there's
• 56:19
  the causality, kind of, you know, that area of, of there was
• 56:27
  also the graph neural networks. There's a few other areas too,
• 56:31
  but these are these are all like different camps inside of of the
• 56:35
  graph space. They don't always necessarily
• 56:37
  talk with each other, but I think it's really fascinating
• 56:40
  now that we're starting to see more and more hybrid
• 56:42
  integrations of them. Yeah, I like to point out the
• 56:48
  fundamentally graph and table are two sided, same coin.
• 56:54
  It is as a physicist like we look at the sound music both
• 57:00
  from frequency domain like is a CDEF, what's the frequency
• 57:05
  distribution and also look at what waveforms like time domain
• 57:10
  like some some situation you want to filter or you want to
• 57:14
  access more on the frequency domain, some sometime makes more
• 57:17
  sense on the waveform domain. The the same data like a like a
• 57:23
  graph essentially is a joint I call like.
• 57:27
  If you think about the large language model neural network,
• 57:31
  it's a graph, but it's a gigantic extremely sparse
• 57:37
  matrix, which is table, right? And the big the fact that
• 57:41
  because like it's a it's a such a giant sparse matrix causing
• 57:46
  today every NVIDIA is really hard because NVIDIA has this
• 57:49
  GPUs that can process those matrix.
• 57:52
  But guess what? And the we might bring consume
• 57:56
  about 19 Watt energy. The, the, the, the, the GPU
• 58:01
  running larger model, large language model consumes 10s of
• 58:04
  thousands of Watt of energy to, to get similar computation
• 58:08
  needs. And that's extremely
• 58:10
  inefficiency. Even though the, the competition
• 58:13
  unit is much smaller than my neuron, you think it should
• 58:18
  supposed to compute a higher efficiency.
• 58:21
  That's precisely because they're dealing with extremely sparse
• 58:24
  matrix. If they're not dealing the
• 58:27
  neural network as an, as a graph, they're dealing neural
• 58:29
  network as a matrix. And that's fundamentally the
• 58:32
  problem for the power efficiency.
• 58:34
  So there are certain certain models that come up that really
• 58:37
  deal with AI as a graph that several order magnitude save in
• 58:42
  the in energy consumption. So in the reward application,
• 58:48
  the the one reason why graph hasn't been taken off as we all
• 58:53
  think for the past 20 years, like, oh, graph, graph going to
• 58:56
  take off. But no, it did not.
• 58:58
  The fundamental problem is because we are so familiar with
• 59:02
  all the all the tools and methodologies like workflows is
• 59:06
  well established in the tabular based way of thinking.
• 59:11
  It's like the the Department of Transportation do not release
• 59:15
  the flight data as a graph, they released as a table.
• 59:20
  Is it easy to access? We have all the toolings that
• 59:22
  mature to change that is extremely difficult.
• 59:27
  So in the way I I would argue that AI is always almost made
• 59:32
  for graph because AI suddenly allow you to process
• 59:36
  unstructured information like emails, reports this from this
• 59:42
  like a podcast, transcriptions, like videos into a structural
• 59:47
  form that computer can access. But guess what, it is a graph
• 59:52
  that AI will convert those data into like.
• 59:56
  So now you suddenly have this, some people argue, I think it's
• 1:00:00
  like 80% of information existing unstructured form.
• 1:00:04
  Some people argue that even that percentage even larger.
• 1:00:07
  So the AI suddenly make this like the majority of the
• 1:00:12
  information available for analytic workflow ad assessment.
• 1:00:18
  And The funny thing is, it needs graph to do that.
• 1:00:23
  So in in the way that I my assessment is because of AI,
• 1:00:28
  because general AI we actually entering a the boom like
• 1:00:33
  exponential growth error of a graph because the availability
• 1:00:38
  in the data. It's like the Internet of
• 1:00:41
  Things. We've been waiting for it to
• 1:00:43
  happen since 20/20/10 or 2005 whenever, and it's always just
• 1:00:50
  around the corner. But now it it does make sense
• 1:00:53
  that if you have all of this unstructured data and you have
• 1:00:56
  these relations, then that sounds like a graph to me, yeah.
• 1:01:02
  And, and going back to like 1980s era AI, hard AI, you know,
• 1:01:07
  whether we're talking about like a star, B star kind of
• 1:01:10
  algorithms or talking about planning systems, all of these
• 1:01:13
  were expressed as graphs. And like, you know, some of the
• 1:01:16
  early thinking that that was like pre Google, that led to
• 1:01:20
  Google, they were talking about graphs.
• 1:01:23
  Some of that work actually came out of like groupware, but based
• 1:01:25
  on graphs. So it's there.
• 1:01:28
  Funny you say that because we had one of the talks at the AI
• 1:01:32
  Quality Conference back last year was from the guy who
• 1:01:37
  created Docker, Solomon, and his whole talk was really like,
• 1:01:41
  everything's a graph. If we really break it down, it's
• 1:01:45
  just it's all graphs and how one thing relates to another thing.
• 1:01:49
  I'll throw I'll throw something else in to kind of go back to
• 1:01:52
  our early part. We were talking about E meets W.
• 1:01:57
  There's a book that a really favorite book, though from early
• 1:02:00
  days. This is like going back to the
• 1:02:03
  early 90s, but early days of neural networks about this, this
• 1:02:08
  idea of like, yeah, there's some conventions in the West.
• 1:02:11
  Maybe we can back off. It's by USC professor called
• 1:02:14
  Bart Costco. It's called fuzzy thinking and
• 1:02:18
  sort of his critique of science, but more from a lens of, of more
• 1:02:22
  Eastern perspectives. I know that this book is like
• 1:02:26
  more than 30 years old, but I, I, I think that there's some
• 1:02:29
  really great perspectives there that weigh in a lot, especially
• 1:02:33
  what we was saying about like where are we now with LLMS and
• 1:02:36
  how we're leveraging this in the context of graphs.
• 1:02:40
  So I think the other thing it was there anything else that you
• 1:02:44
  guys wanted to talk about before we jump?
• 1:02:48
  I I know there's there's a lot of cool data visualization stuff
• 1:02:52
  that you're doing way. Yeah, I just want to add one
• 1:02:55
  thing. I just want to say the
• 1:02:58
  visualization is not the end. The goal is to support
• 1:03:05
  analytics. So I, I, I, I know everybody
• 1:03:09
  when it comes to the graph, talk about graph visualizations.
• 1:03:13
  But in my mind, it's what's really is what we need is visual
• 1:03:17
  analytics. How can we visually transformed
• 1:03:21
  informations? How can we visually go from like
• 1:03:24
  information that was suited for data management, for data
• 1:03:28
  capture that that was so you can access work them step by step
• 1:03:33
  towards information that's suitable for presentation, for
• 1:03:37
  answering the specific questions in that particular domain.
• 1:03:41
  So that steps requires a transformation of data is not
• 1:03:45
  just like a filter, but also fundamentally in the graph
• 1:03:50
  schema mutation like the schema you have for the data capturing
• 1:03:56
  is not a schema suitable for presentation.
• 1:03:58
  There are two different things is that if you think about in
• 1:04:01
  the big data era, yeah, the the the developer of the MapReduce
• 1:04:06
  allow you to have this like step by step flow of information from
• 1:04:12
  the original captured tabular format into final very different
• 1:04:18
  table that you can present in graph is the same thing that
• 1:04:23
  that the the graph analytic needs is a step by step like we
• 1:04:28
  call it calculus or operators to transform your data from from
• 1:04:32
  the the form that being captured to the form that you want to
• 1:04:36
  present. Or to answer the question now
• 1:04:39
  that calculus is best done yet. I, I, I, I think it needs needs
• 1:04:45
  to be in two forms. It needs to be in the form that
• 1:04:48
  you can process data in large quantity, like a large graph
• 1:04:54
  step by step mutates, but also needs to be visually.
• 1:04:59
  You need the same same set of a parallel set operator that a
• 1:05:04
  data analyst, but ideally a domain expert, not a not a data,
• 1:05:11
  not somebody who can write Python or cipher queries or GQL,
• 1:05:18
  but somebody with the domain knowledge and look at it because
• 1:05:20
  graph's so visual, you're like, hey, I want to simplify this.
• 1:05:24
  Oh I know Paco and Wei has so many meeting points.
• 1:05:30
  Let's abstract that out. Let's just create a single
• 1:05:33
  reading stream that way inference like Wei and Paco that
• 1:05:37
  they know each other and get rid of all the other informations.
• 1:05:41
  So, so this all maybe say, hey, Paco knows like a meeting people
• 1:05:46
  maybe I underestimate a little bit of Paco, so sorry about
• 1:05:49
  that, but. No kidding you.
• 1:05:50
  Probably know more than that, but let's let's like from the
• 1:05:53
  graph, we can quickly compute this number and put it in the
• 1:05:55
  pocket, make Paco very, very big because Paco knows a million
• 1:05:59
  people, right. So that kind of operation is
• 1:06:03
  highly intuitive. So I want to stress this, the
• 1:06:07
  visualization for graph is not end.
• 1:06:10
  The visualization for graph is tool you use to transform the
• 1:06:14
  graph. Yeah, to get you the answer.
• 1:06:16
  That's a way point. Very good.
• 1:06:18
  Yeah, it yeah, that is very in line with what you were saying
• 1:06:23
  earlier on how when you don't know the question, that's
• 1:06:29
  sometimes the hardest part. And so being able to wrestle
• 1:06:32
  with the data in different forms, one being the visualizing
• 1:06:37
  it in different ways, that's one tool to hopefully help you get
• 1:06:42
  to the answer or first step, the question, which can then lead to
• 1:06:48
  the answer you're looking for. Yeah.
• 1:06:50
  And to mutate the graph visually.
• 1:06:53
  So you can, you can, you can. You can start poking it, yeah.
• 1:06:58
  Yeah, exactly it. It does feel like the ability to
• 1:07:04
  just mutate the graph is such a strong tool because of all these
• 1:07:12
  different reasons that we had mentioned when it comes to the,
• 1:07:16
  the depth and the way that you're able to look at the
• 1:07:19
  domains or you're able to just find anomalies or find different
• 1:07:25
  data quality issues, whatever it may be, whatever your use case
• 1:07:28
  is, it, it's very cool. I It does sound, though
• 1:07:32
  instinctively, a bit manual though.
• 1:07:34
  Right. So far, I, I, I think way has
• 1:07:40
  brilliant examples what they're doing like with side XR of
• 1:07:43
  leveraging 3D visualization, zoom in, zoom out in conjunction
• 1:07:47
  with algorithmic ways, using graph algorithms to sort of
• 1:07:51
  focus the lens, focus the Searchlight.
• 1:07:55
  I, I think that more can be automated over time.
• 1:07:58
  And maybe this is where agents come in is actually helping
• 1:08:01
  determine how to how to be the cinematographer there on the
• 1:08:05
  graph. Yeah, so there's a definitely a
• 1:08:08
  way of helping you to look perspectives and is very often
• 1:08:12
  we deal with the data that's both graph connect nature, but
• 1:08:15
  it's also dimensional like yeah, each nodes has so many like a
• 1:08:20
  properties, each property is a dimension.
• 1:08:22
  So it's high dimension information.
• 1:08:25
  So which dimension set you want to take in combination with the
• 1:08:30
  network information to help you to see be able to have a
• 1:08:35
  versatile way, flexible way of to choose the dimension set.
• 1:08:39
  Or is a very often like when you when you shift from 1 dimension
• 1:08:42
  to the other dimension, you review some flockings of things
• 1:08:46
  going together, some clustering start happening areas that hey,
• 1:08:49
  those things always move in the same direction.
• 1:08:52
  So those signals help you to formulate a lot of ideas,
• 1:08:56
  instincts from the data. And then when you see that
• 1:09:00
  information, you would next thing you want to know, hey, I
• 1:09:02
  want to capture that as a feature.
• 1:09:05
  Now can you represent that as a feature to that become what you
• 1:09:11
  see become a thing that become an entity in your in your
• 1:09:15
  visualization that you can yeah, I can put back in there.
• 1:09:20
  And that is the visual analytics.
• 1:09:25
  Whoa. So capturing it as a feature and
• 1:09:28
  then you can feed it into the tabular data in a way.
• 1:09:32
  Yes, exactly. Guys, this is awesome.
• 1:09:36
  Is there anything else that you want to hit on before we stop?
• 1:09:40
  I feel like I've learned a ton just from talking to you all.
• 1:09:43
  I knew it was going to be a great conversation.
• 1:09:45
  I was hanging on to my seat this whole time like Oh my God I'm
• 1:09:49
  learning to. Yeah, in terms of the cross
• 1:09:53
  domain, I, I want to share 1 funny examples like how
• 1:09:57
  difficult the cross domain is. So, so in this example is
• 1:10:00
  extreme cross domain. So, so I organize a kinetic
• 1:10:06
  arts, dance and science like like nonprofit.
• 1:10:12
  So one thing we do is every week, every Wednesday, we bring
• 1:10:16
  people in the engineers science domain and people in the dance,
• 1:10:21
  art, music domain together. We explore something together
• 1:10:24
  and have a conversation. The very first meeting we when
• 1:10:29
  we bring people together that happened about like 11 years
• 1:10:33
  ago, we had about 20 people sitting in the room.
• 1:10:37
  Everybody like a very vibrant conversation.
• 1:10:40
  And then that's the sudden realize something that it is
• 1:10:43
  true that everybody speak English, but nobody can
• 1:10:47
  understand each other because they're using SIM vocabularies.
• 1:10:55
  But because of domain, just like Paco talk about earlier in the
• 1:10:59
  in the enterprise setting, because of the domain
• 1:11:02
  difference, they means totally different things.
• 1:11:06
  A physicist that talk about energy, we have very concrete
• 1:11:10
  things that we call energy. A dancer called energy is a very
• 1:11:16
  different way of energy. Yeah.
• 1:11:19
  When when the computer people talk about Python, we're not
• 1:11:24
  talk, we're not talk about a snake.
• 1:11:27
  But the dancer, when they hear Python, they're like, why you
• 1:11:30
  bring the snake to the conversation?
• 1:11:35
  So I think in the just accurate what Paco said earlier in the in
• 1:11:39
  the enterprise data context, that domain is very, very
• 1:11:46
  important to be aware of the domain.
• 1:11:48
  Like knowing the limit of the domain and how to find a way to
• 1:11:52
  cross domain for for us is generally a lot of compensation.
• 1:11:57
  I think it's a human problem. It's not a technical problem or
• 1:12:01
  techno can help but all into that much.
• 1:12:06
  We had a conversation on here a few months ago with folks who
• 1:12:11
  had created a data analyst agent and they said one of the hardest
• 1:12:16
  parts for the success of this agent was to 1st create a
• 1:12:24
  glossary of business terms. So that the agent and, and
• 1:12:29
  really trying to nail down these fuzzy words and these words that
• 1:12:37
  maybe for one person they mean one thing and another person
• 1:12:40
  they mean another thing. And the quintessential example
• 1:12:44
  of this is an MQL. When you're at one company and
• 1:12:49
  MQL or when you're on one team and MQL is one thing.
• 1:12:52
  And when you go to another team and MQL is another thing, they
• 1:12:57
  all mean marketing qualified lead.
• 1:12:59
  But when does that person become a marketing qualified lead?
• 1:13:03
  What do they have to have done or what stage are they in?
• 1:13:08
  And so the agents may understand and the LLMS understand what an
• 1:13:14
  MQL is kind of, but you really have to flesh out this glossary
• 1:13:19
  to let them know all of these different terms that you use and
• 1:13:26
  that are in your database. So when the agent needs to go
• 1:13:31
  and pull how many MQLS did we have last week, it understands
• 1:13:36
  what that means. Yeah, that, that's your semantic
• 1:13:39
  layer right there. That's your that.
• 1:13:40
  That's a controlled vocabulary that you know, you put enough
• 1:13:43
  these together, you get your ontology.
• 1:13:45
  Yeah, yeah, yeah, exactly.