923: Graph Algorithms, GraphRAG and Causal Graphs, with Graph Guru Amy Hodler

26:11Jon Krohn:Wow. Those were a lot of great examples that you were able to just enumerate there. I love that the telomere one is particularly interesting to me. I have a PhD in neuroscience. Oh, fascinating. Yeah. And so these kinds of things around. So some listeners may be aware that your lifespan as an organism is basically dictated by how long your telomeres are, which are things that are made by, I think basically always your mom and your dad when they produced the the sperm and or the egg that led to that led to creating you the the telomeres were extended um basically the it's it adds kind of garbage dna or dna that doesn't have any genes in it on the ends of your dna so that over your lifespan every time your cells um replicate so you go you You go from a sperm and an egg to a single cell, and then two, and then four, and then eight, and 16.

27:15Jon Krohn:And every time that happens, your DNA has to be copied. And then throughout your lifespan, as your cells continue to divide, your telomeres shrink, shrink, shrink, because the mechanism that copies your DNA needs something to clamp onto, and so it clamps on right at the end. And so those telomeres that were added on by your parents, thank them for that, is gradually eaten into over the course of your lifespan. And so a lot of what we consider to be aging is your telomeres being shortened. And your telomeres run out in some of the tissues in your body, and you're starting to eat into real meaningful genes.

27:58Jon Krohn:And so people who want to solve aging, a lot of it is about figuring out how you can have your telomeres grow while you're still alive without needing to create a sperm or an egg to be able to have those telomeres extend. And the tricky thing is that telomeres extending is also one of the key causes of a metastatic cancer tumor. Yeah, it's definitely a balance. Yeah. I mean, and the graphy element of that is just understanding the connections between things. And biology and biologists are very open to and understand a graphy way of looking at things. So there's a lot of use cases in biology as well.

28:44probably not where the biggest funding is in graphs, but, but there's just a natural understanding when you, when you talk to people with any kind of a healthcare or life sciences background that they just kind of, they get the fact that things are connected and the, the, you can use graphs to understand the flow of resources, which you can model sugar, you know, the flow of sugar through the body in that way. So there's, you know, if any, Anytime you're looking at highly complex systems, yeah, you end up getting into graphy theory behind that.

29:16Jon Krohn:You mentioned there that biology maybe isn't the place that a lot of resources are going for graph research. Where do you see a lot of resources going? If our listeners are thinking, wow, graph sounds super interesting, or I already am a graph expert, I'd love to find lots of fertile ground, where do you think they could find it? It's changed over time. So I would say, I hate to say fraud, fraud, and fraud, but finding bad actors. And when I say resources, I'm thinking commercial funding. And if you were doing a startup, what area would you focus in? So fraud is always, I want to say evergreen, but it is crowded.

29:55So you have to have a different way of looking at it or some kind of unique quality of how you're looking at it. So, you know, always throw fraud in the mix, whatever you're doing, just throw it in there and it's a door opener and people understand it. As cybersecurity is hot, again, it's starting to be a little crowded, but it's hot for graphs because, again, talking about how people try to hide and just the increase, the exponential increase in cybercrime right now means that people are looking at all different ways of tackling it. And so I think graphs and cybersecurity make a lot of sense, and in particular, when combined with other technologies as well.

30:39So it's another view on what you might already be doing. There's also some emerging use cases that I don't know if I would say reemerging that people had talked about previously, but are now becoming more tractable because they're combining it with LLMs and Gen AI. one of the ones that I would say are fairly low hanging fruit and probably people could build, I don't want to say build on their own, but for lack of a better word, you could probably build your own is documentation analysis. So if you think about like legal documentation, this is one area that I'm surprised not everybody's doing because it doesn't seem like that heavy of a lift, But legal documentation tends to have a lot of links to other caveats, other laws, sub clauses, sub clauses of sub clauses.

31:32And if you change this, what's the ripple effect of the connections to something else that might, you know, you might actually be signing something that is impossible to fulfill because of the rippling connections and requirements. And so looking at legal documentation and helping your lawyers and your business analysts understand it better is easy, low-hanging fruit. It's easier now because you can use the LLMs, the parse, the language, and then use the graph to understand the connections between the different documents. Just for basic document review and understanding, we even see this in help desks.

32:10So helpdesk looking at their own documentation, their own complicated documentation. And I talked to somebody just in April that was trying to figure out their massive reams of documentation and help their support line get through their own documentation because there was just so many different connections and links and threads. Anytime you want to pull a thread, but also lawyers as well are looking at it. So those are some of the interesting use cases that aren't necessarily new, but because of Gen AI coming in and large language models coming in, the combination are allowing these old new use cases to become just easier to accomplish.

32:53Jon Krohn:So those kinds of use cases that you were just describing there, like the legal document search, is that GraphRag? It can be. It can either be straight graph or you can use GraphRag. I would say right now, everybody's excited about GraphRag. So I would expect that to be somewhere in the mix, but it doesn't have to be, and it wasn't always. So Caterpillar uses Graphs and has for years for some of their support documentation and support staff. And that was before any of us had talked about RAC. But I would imagine, I don't know, haven't talked to them lately, but I would imagine And now they've got RAG in the mix and they're probably have a chat bot that, you know, is talking to that documentation as well.

33:35Jon Krohn:That's right. So let's talk about that in more detail, I suppose, if this is such an exciting area, GraphRAG. So retrieval augmented generation RAG, probably a lot of our listeners out there are already familiar with the term, but it's this idea of being able to search over. Well, so you start off by having, say, a large number of legal documents or Caterpillar's uh in caterpillar that's like the it's not a tractor company what do you call those i would call it heavy heavy equipment because they do a lot they do a lot of different equipment yeah if like like a lot of equipment with tracks on them it's the caterpillar right uh yes yes yes lots of lots of legs lots of uh points on the ground i suppose yeah um so yeah so whatever these use cases, you have lots of documents.

34:24Jon Krohn:You start by converting all of those documents into a vector representation, meaning a location in some high dimensional space. So you can visually think about it like in a two dimensional space or a three dimensional space, but in practice, it might have hundreds or even thousands of dimensions and the documents end up closer together because they're more, the meaning in the documents is more similar. And so LLMs have been great at figuring out that similarity and being able to get related documents close to each other. And so once you have all those documents close to each other, you can use reg, retrieval augmented generation, to say have a user ask a question.

35:05Jon Krohn:And that question is then in real time very quickly turned into that same kind of vector representation, the same kind of location in a high dimensional space. You can find related documents, you can pull them back, And then you can use an LLM to search over that relatively small number of documents you pull back to generate some response to the user's query. So, I mean, you can correct or change anything that I just said about RAG. But I'd love to hear now how we can take that RAG idea and how it becomes graph RAG. Yeah. Yeah. So there are a number of reasons why RAG has, some people feel like we're past the RAG moment.

35:51It's just things get hot, people get disappointed, and then they want to improve it. But there are several reasons why RAG has plateaued, I would say, a little bit in its performance. And that's why people are bringing in Graph as well to help out with it. It doesn't replace. It augments. Augments? RAG? Anyhow, I don't know what the acronym of GRAG or who knows what it might be. But you've seen problems with RAG basically with things like getting a plateau when your answers and your questions, your questions and answers don't align very well, especially semantically. And so you have this kind of meaning gap.

36:40So semantic just means meaning. But you have this meaning gap between what somebody's asking and the question, and the context isn't really well understood. So you see things like that. You see things like the ability to diversify the data you're pulling from. So RAG does really well with trying to extend what your retrieval is already doing. And so the retrieval effectiveness, you know, pure semantic vector search can be insufficient. So you might get an approximate match rather than an exact match, or you might want a match that has topological significance that you're just not getting. You might also need to do multi-step reasoning.

37:27So if you think about agentic rag where you need larger context windows, or you need to pull in diverse data sets, or you need to look at the previous responses to come up with a better response, graph can help with that in general, because again, we're looking at relationships, and that gives us context, and we're able to pull in topology as well. And so I think of graph and And graph rag gives us, vector search gives us really good summarization. It gives us really good fuzzy matches. If you're doing a full text search, you can get some exact matches. But graph gives us structure as well. Again, it's that topology.

38:18And that gives us matches and answers of things like dependencies, serial, things that might have serial dependencies or you might have to do them in a sequence. It's also good at aggregations, you know, especially if you're doing aggregations over multiple data sources and things like that. So I always think of, I think of graph in the graph rag as rag and. It's like you, you know, just, you don't want to just have one type of, you know, augmented generation. You want to include those other capabilities. Graph lets us do that. And it also helps us stitch things together because of the connections it has as well.

38:59Jon Krohn:So am I correct in understanding that they kind of happen in parallel, that you would be doing like a vector search as well as a graph search in parallel, and so then you get the best of both worlds, you get the fuzziness of the vector search with the specificity of a graph search? Yes, actually, when you think about vector and graph in graph reg, or it really should be called hybrid rag. That is the very thing you do see. And the results that I've seen, and this is an evolving space. So ask me in six months, question might be different or the answer might be different. But what I've seen is this hybrid in parallel that seems to get the best results.

39:39And then you have a method to then blend the results at the end and then augment that in an ongoing fashion.

39:48Jon Krohn:Very cool. So I'm sure a lot of our listeners out there are saying very cool with me right now. And so all of our hands-on practitioner listeners, whether they're data scientists or software developers or AI engineers, they might be wondering how they should be getting started with graphs. It could be GraphRag or just getting going on graphs in general. What are the key tools out there? Or I guess, what are the ones that you recommend people get started with? Yeah, I always think about thinking about your use case in mind. I mean, graphs are just fun in and of themselves. There's a lot of material.

40:23If you Google graph theory, there's a lot of fun material. But I would think about your use case and applying it in that manner. I try not to recommend specific tools, but there are a couple vendors that have really nice one-on-one material that you can use to get up to speed. I would say right now, don't lock yourself into one methodology. In the past, several years ago, like maybe three, four years ago, it was assumed if you were going to add graph to your capabilities, you had to have a graph database. That is no longer the case. There are several vendors out there that allow you to project a graph with a computational graph, so a classic graph engine or a graph layer.

41:09There are several vendors that also allow you to poke into your table data and ask a graphic question of your table data and just use that projection to answer the question and then drop it. And so depending on your need, there's a lot of different options for you. And there's no one option that's good for every situation. So looking at your own tech stack is probably real important. There's a couple conferences I really like. If anybody, to give a shout out to the Open Data Science Conference group, they do conferences in the East Coast, West Coast, Europe. They also do virtual conferences or virtual training as well.

41:57I like that because you get a broader spectrum of opinions and different ways of looking at graph, but also how it fits into the bigger picture as well.

42:08Jon Krohn:ODSC is my favorite conference as well. And I'm not just saying that because they sponsor the show. And they literally, they actually, in your episode, Amy, you can possibly know this because you don't know what ads we're going to put in. But at the, around the 30 minute mark in this very episode, there's an ODSC West 2025, uh, sponsor message. Uh, I absolutely love, uh, ODSC. I think it's the best for hands-on practitioners. Uh, because I'm in New York, I have the privilege of, uh, speaking at East in Boston in the spring most years. Um, and I get out to West, uh, or it's pretty much always on Halloween.

42:51Jon Krohn:uh i i get out there whenever i can sometimes my travel schedule doesn't permit it but if it does i'm always at west as well i love it michelle and i will be there this year i'll have to try to make sure i get there we could do something fun yeah we could do something we could do stand-up comedy um you should just stick to technical stuff you'll just terrify me I don't think I'm never funny on purpose, but I'm often funny. So there you go. I actually, I recently was on another podcast called the Modern CTO Podcast, which is a cool show. And the host is very funny. And he actually, he now spends other than, so he professionally hosts the Modern CTO Podcast.

43:40That's his main job.

43:42Jon Krohn:but his a huge amount of his time uh approaching uh a full-time job is being a stand-up comedian now as well uh so joel beasley uh and it's pretty interesting to hear i think by the time your episode is out my appearance on the modern cto podcast should be out so there's cool things that he talks about in the episode around using data in analytics to and using llms to uh to review all of his standup routines. And so he's trying to get a particular number of laughs per minute from the audience. And he's trying to match. So he's done analytics around what other, you know, professional standup comedians, if you get a Netflix special, how many laughs do you get per minute on average?

44:27Jon Krohn:And so he's trying to work his way up to that. Wow. That's terribly, that reminds me of a story. I had somebody once asked if I would come to their offsite, They were quants. And they're like, can you come to our offsite, talk about graphs and be funny? I was like, you're terrified. I don't. OK, we'll see. So that's a lot of pressure, a lot of pressure when you feel like you have to be. So, yes, you were very diplomatic about your answer, the tools, which I greatly appreciate. And I suspect that that's somewhat related to you being founder and executive director of Graph Geeks. tell us about graph geeks uh how it got started and why someone should reach out to you uh if they need support yeah so graph geeks i love to tell people that i started it because i got lonely but that is actually part part part of the truth so i'd been in the graph space for several years um i don't know over 10 i guess and i have been at several different vendors and when you when When you're working for a vendor, it's wonderful to give you focus, but you also have that focus also gives you a particular way of looking at the graphs landscape and world.

45:46And so when I left the last graph company I was at, I just had a lot of wonderful conversations with people that were either starting things up or had a lesser known, lesser funded approach to graphs and really wanted to highlight that. And the other thing is I was just missing my graphy friends. So the interesting thing about graph folks, they are often a small subset at a company. You usually don't have a 100-person graph team. You sometimes have a one - or five-person graph team. And so we hold on to each other pretty closely. We talk to each other just naturally when we look at the world through relationships.

46:25So we naturally keep our relationships. We help each other out over the years, and we stay in contact. We go to the same conferences. So it's a bit of a tight-knit community. community. And so when I left the vendor space, I missed immediately missed my community and how we looked at things. And I was looking around and I realized there was no vendor agnostic graph community online that I could easily reach out with and interact with on almost a daily basis if I wanted to. So I thought, what the hell? I can start this. If I start this and I fail, nobody can fire me. Why not? And so I started GraphGeeks.

47:03We are graphgeeks.org. If people are looking for us, there is a Discord, a sizable Discord community that helps each other out. So we have a I need help section. We have resources, probably not as many training resources that I would like to get right now, but there are a lot of graph practitioners. So if you're stuck in the middle of a really gnarly graph problem and you're looking for help, you can go on the Discord channel and say, hey, I'm stuck. What have you seen? If you've read a paper, love it, hate it, don't understand it, reach out. If there's somebody you would like to see on the Graph Geeks, either webinar, podcast, what have you, I've got a YouTube channel.

47:49Let me know. I try to get that information out there and help people with things on like, what is the difference between RDF and property graph? We've got a couple hours of that material out there. How do you design to optimize a query? We've got some stuff on that, you know, so there's very, there's a lot of interesting topic, but it's the well is deep. And so, you know, over time, we'll just be adding more and more material. I also have many volunteers and opportunities for volunteers. So if people just want to like geek out, we do that. Yeah. So that's, that's how I got started.

48:26Jon Krohn:Nice work, Amy. It sounds like you're doing an amazing thing for the graph community. You are now an invaluable node in that network. A pivotal node. Call me a pivotal node. A pivotal node. Thank you. So before I let you go, one last technical question that I want to get some insight from you on is what is changing in graphs? You know, what's next? So, you know, we've spent this episode learning about why graphs are cool, what they're useful for. You gave us some direction on tools that we could be grabbing. And so, yeah, like what's next? Some of the things that you mentioned to me before we started recording included multimodal, included graphs for LLM memory and causal graphs.

49:12Jon Krohn:Maybe we could touch on each of those quickly. Yeah, so I'll just, I'll quickly go through the major changes. One is that I already discussed a little bit is framework diversity. So there are like the query engines are getting better. So you don't have to have a database. Different types of graph databases are becoming available. You also have hyperscalers that are getting into reentering the graph space. So lots of choices on framework. So that's a big one. Multimodal, I would put out, well, maybe I should say graphs and AI and what bringing them together is allowing from a use case standpoint.

49:54We talked a bit about that. And then multimodal, which is being able to graph different types of data. So one of the things a colleague of mine, David Hughes, shout out to him, and I do present on is this idea of modeling an image as a graph. And so most of the time we talk about graphs, people think about lexical graphs, so graphs of words or graphs of concepts. Those are the traditional uses. However, you can graph an image. So if we have a picture of me holding my coffee cup, you have the main image is Amy, but there's a coffee cup in front of me to the right. And that relationship has meaning as well.

50:40And so being able to connect those as meaning allows us to do things if we're looking at, for example, and we've done this, looking at a ship, a fleet of ships, and some are ahead of the other. And you can graph that relationship. And then if you look at that relationship over time, you can also estimate the speed. Are those ships coming together? Are they pulling apart? Do they look like they might be antagonistic to each other? So there's all of these things that you can do with different data types. So again, moving to images, we've also added in audio to that. And so for example, we did that with police cars, where you hear them in a video frame, but you don't see them.

51:25And with Doppler effect, you can tell what direction the police cars are heading. And you can do that by graphing it. And to me, that's exciting, not just from a graph rag standpoint, which is what most people want to talk about, like how do I use that with my graph rag, but just this idea of something we have done with graphs forever, which is modeling the relationships between things, we haven't extended it to things in a image, or things in audio. And to me, that just opens up to all sorts of other use cases, like detecting things in sonar, to, again, directional speed in an image, to understand a whole, a grouping in an image of people.

52:10Is there a relationship that we can infer based on how people are standing next to each other? So there's that to me, sorry, multimodal, very, very fascinating area, really cool. But the other one that, or the other two that I would be remiss if I do not mention them first is graph as memory. So graph provides us a way to capture context, and context is really important for AI. And so if you think about the context windows of an agent, they're relatively short right now. So there's a couple of really interesting papers, ZEP, which I have sitting on my desk right now, temporal knowledge graph architecture for agent memory, a must read if you're interested in extending agent memory.

52:57And then MEM0, building production-ready AI agents with scalable long-term memory. Those two papers, really significant, I think, in looking at how you use the context saving ability of a graph to store memory for agents, either for just very simply extending the context window. And you can basically store context and then retrieve it later when you need it or even longer memory. So going beyond a typical context window. That I think is gonna be super hot by the end of the year. If you're into graphs and you haven't thought about graphs as memory for agents, take a look because that's something that I think in six months or less people are gonna be talking about.

53:47Jon Krohn:It is something that's been on my radar as well. That MEM0 paper is something that keeps coming up. So yeah, I agree. Keep an eye on that. The final thing as far as emerging, if we want to add this at some point, is causal graphs. So understanding not just the prediction, being able to predict that something happened, but why it happened, I think is something that in two-ish years will be the next hot thing because we're doing a really good job of predicting the next item in a sequence, but we don't know why that sequence happens. and graphs because you have those two nodes, those circles with an arrow between it, those arrows can, or those links can have, can be an arrow.

54:32And so you can link things in a sequence and you can try to understand influential cause, um, in, you know, whether you're talking about economics or biology or, or what have you. So that's the other area to look out for, but probably a little further ahead.

54:46Jon Krohn:Nice. Thank you for that. Look into the future, even though things move quickly. It does feel like your insights, you know, looking at six, 12 months, maybe even longer are going to end up being really, really helpful. Uh, Amy, this episode has been incredible. You're an absolute pro. It's unsurprising that you host your own podcast, uh, that you do all of this, uh, public facing graph, uh, knowledge distribution work because people, you wouldn't, people wouldn't be able to tell us because we edit all the episodes, but everything in this episode was done in a single take. And so there's been no retakes.

55:24There's been no kind of umming and ahhing,

55:27Jon Krohn:thinking about answers. Amy's just done everything off the cuff flawlessly. It's been a joy recording with you before I let you go. I ask all of my guests for a book recommendation. Do you have one for us? Yes, actually I joined a book club, which I highly recommend joining an old fashioned book club. it's really wonderful to talk to humans about their, what they are thinking. The book is called These Strange New Minds. And it is about LLMs and thinking about them from everything from a technical standpoint to a philosophical standpoint. Are they thinking? Are they not thinking? Do they have the appearance of thinking?

56:05What does thinking mean? And what do we want LLMs to say? How do we look at optimizing and regulating? So I think in this moment in time that we are all in, understanding LLMs a little deeper and from a couple different viewpoints is just a huge benefit. And I highly recommend the book. It's exceptionally well-written.

56:30Jon Krohn:So it's nonfiction. I mean, it's not like a fictionalized... Yeah, yeah. No, no, it is not. It is a multidimensional look at large language models and what it means for models to actually speak back to us. And so this book club that you're in, is this the typical kind of book that you read? Are other people in your book club also kind of data AI people like you are? What's going on here? Yeah. Yes, it is a typical book. We also did one on responsible AI called AI Snake Oil. That's another really good one to look at, but definitely has its opinion. So far, yeah, we've been picking data and AI topics.

57:13And it's a group of us that are in the tech industry. But it's even within the tech industry, these concepts are so large. Having people with different viewpoints, we have somebody from a security background. We have somebody with a responsible AI background, an ethics background. We have people that are just building product. So I think having diverse people and mindsets, people from different countries as well, really shows that even when we have our own opinions on something, there's probably a different way we haven't looked at it. So anyhow, join a book club. Doesn't matter what it's for.

57:53Jon Krohn:It's a great idea. It's something that I, it's on my list of things to do. I hope to get there. I even just reading more, it's something that, yeah, I just, I fantasize just like my fantasy about being under leaves. I also fantasize about being under the leaves of a book. Yeah. Yeah. Yeah. We can talk about that another time because I do think about doing more. It has been very rewarding and it's very useful. So that's why we do these book recommendations at the end of every episode. I know how valuable it is. Even if I would love to read every single book recommendation that I get from my guests, and there's just no possible way I can do it.

58:33Jon Krohn:Fantastic, Amy. So yeah, as I already said a few minutes ago, this has been an amazing episode. You've been an extraordinary communicator. We've already talked about Graph Geeks, which people can visit at graphgeeks.org. Where else should people follow you after this episode to get more of your brilliant thoughts? Well, they can also follow me on LinkedIn. I'm very easy to find. I have an unusual last name. So Amy Hodler, H-O-D-L-E-R, easy to find me there. That's primarily where I post that in I'm on Graph Geeks. Fantastic. Thank you so much, Amy. It has been such a joy to have you on the show.

59:08Jon Krohn:And yeah, hopefully we can get you on again in a few years and see how the graph world has come along. Absolutely. Thank you, John. It's been a real pleasure and I've had a lot of fun and I appreciate all your really insightful questions.

59:22Jon Krohn:Wow, another fun and informative episode today. In it, Amy Hodler covered how graphs capture relationships in data using nodes and edges with properties that include quantities, strengths, and various data types. How graph algorithms like PageRank compute over network topology to reveal insights about structure and behavior that traditional statistical methods miss, making them powerful for applications like fraud detection, supply chain optimization, and recommendation systems. She talked about GraphRag and how it enhances traditional retrieval augmented generation by adding structural context and topology to semantic vector search.

59:57Jon Krohn:And she talked about exciting emerging graph applications, including multimodal graphs, graphs as memory systems for AI agents and causal graphs. As always, you can get all the show notes, including the transcript for this episode, the video recording, any materials mentioned on the show, the URLs for Amy's social media profiles, as well as my own at superdatascience.com slash 923. three. Thanks to everyone on the super data science podcast team, our podcast manager, Sonia Breivich, media editor, Mario Pombo partnerships manager, Natalie Jaisky, our researcher, Serge Massis, writer, Dr. Zara Karche and our founder, Kira Larimenko.

1:00:33Jon Krohn:Thanks to all of them for producing yet another excellent episode for us today for enabling that super team to create this free podcast for you. I encourage you to check out our sponsors. You can support this show by clicking on our sponsors links in the show notes. And if you are interested in sponsoring an episode yourself, you can get the details on how at johnkrone.com slash podcast. Otherwise, share this episode with folks who love it. Review the episode on your favorite podcasting platform or YouTube or wherever you consume podcasts. Subscribe if you're not a subscriber, but most importantly, just keep on listening.

1:01:11Jon Krohn:I'm so grateful to have you listening and I hope I can continue to make episodes you love for years and years to come. Till next time, keep on rocking it out there. And I'm looking forward to enjoying another round of the Super Data Science Podcast with you very soon.