Support the show
Video version available on YouTube
Recorded on Sep 27, 2023 in München, Germany
From supply chain to large language models and back
Ishansh realized the potential of data when he was just 10 years old, during his time as a junior cricket player.
His journey led him to ask questions about the mechanisms behind the observed events.
Can large language models (LLMs) help in building an industrial causal graph?
What inspires stakeholders to share their knowledge and which causal discovery algorithms have been most effective for Ishansh's supply chain use case?
Hear the insights from one of the BMW Group's fastest-rising young data science talents.
Ready?
About The Guest
Ishansh Gupta is a Lead Data Scientist at BMW Group. Previously, he worked for several companies, including a legendary German sports club SV Werder Bremen. He studied Computer Science, and co-founded an educational startup during his study years. He has supervised or supported students in various universities, including the Munich-based TUM and MIT.
Connect with Ishansh:
- Ishan
Should we build the Causal Experts Network?
Share your thoughts in the survey
Listen on: Apple Podcasts Spotify
All Business. No Boundaries. The DHL Supply Chain PodcastWelcome to All Business. No Boundaries, a collection of supply chain stories by DHL...
Listen on: Apple Podcasts Spotify
Causal Bandits Podcast
Causal AI || Causal Machine Learning || Causal Inference & Discovery
Web: https://causalbanditspodcast.com
Connect on LinkedIn: https://www.linkedin.com/in/aleksandermolak/
Join Causal Python Weekly: https://causalpython.io
The Causal Book: https://amzn.to/3QhsRz4
Support the show
Video version available on YouTube
Recorded on Sep 27, 2023 in München, Germany
From supply chain to large language models and back
Ishansh realized the potential of data when he was just 10 years old, during his time as a junior cricket player.
His journey led him to ask questions about the mechanisms behind the observed events.
Can large language models (LLMs) help in building an industrial causal graph?
What inspires stakeholders to share their knowledge and which causal discovery algorithms have been most effective for Ishansh's supply chain use case?
Hear the insights from one of the BMW Group's fastest-rising young data science talents.
Ready?
About The Guest
Ishansh Gupta is a Lead Data Scientist at BMW Group. Previously, he worked for several companies, including a legendary German sports club SV Werder Bremen. He studied Computer Science, and co-founded an educational startup during his study years. He has supervised or supported students in various universities, including the Munich-based TUM and MIT.
Connect with Ishansh:
- Ishan
Should we build the Causal Experts Network?
Share your thoughts in the survey
Listen on: Apple Podcasts Spotify
All Business. No Boundaries. The DHL Supply Chain PodcastWelcome to All Business. No Boundaries, a collection of supply chain stories by DHL...
Listen on: Apple Podcasts Spotify
Causal Bandits Podcast
Causal AI || Causal Machine Learning || Causal Inference & Discovery
Web: https://causalbanditspodcast.com
Connect on LinkedIn: https://www.linkedin.com/in/aleksandermolak/
Join Causal Python Weekly: https://causalpython.io
The Causal Book: https://amzn.to/3QhsRz4
010 - CB008 - Ishansh Gupta - Audio
Ishansh Gupta: And what I've personally felt is when I've talked to the stakeholders, it could be also the business experts, the data scientists. One thing everybody says is that it just makes sense. A lot of people are always scared when we talk about causally and all, because it's just so good. So there'll be, are you replacing my job?
What I wanted to do is that maybe I use LLMs to see that how the external world is. Can affect some of the suppliers, but the result I got was really Hey Causal Bandits, welcome to the Causal Bandits podcast. The best podcast on causality and machine learning on the internet. Today we're traveling back to Munich to meet our guest.
He realized the potential of data at the age of 10 while being a cricket player. He's a passionate traveler who visited 25 countries and lived in five. He loves Polish pierogi and a good time at the gym makes him happy. Lead data scientist at BMW Group and a truly entrepreneurial spirit. Ladies and gentlemen, please welcome Mr.
Ishansh Gupta. Let me pass it to your host, Alex Molak.
Alex: Ladies and gentlemen, please welcome Ishansh Gupta.
Ishansh Gupta: Thanks a lot for this warm introduction and happy to be here.
Alex: Happy to have you here. Ishansh, where are we today?
Ishansh Gupta: So we are in this beautiful city in Germany, Munich, in the Bavarian area. And it happens to be a beautiful, lovely day.
Sunny and warm, and especially at this time of the year to have this, it's always exciting and beautiful.
Alex: How did you end up living in this beautiful city?
Ishansh Gupta: It's a long story. I've always been wanting to live abroad. Started with my fascination of wanting to be in U. S. Then, during my bachelor's in India, I got a chance to travel to Cairo for an internship as a website developer.
I really liked the international experience of being abroad, and I knew that this is what I want to do. And even at that time, I really wanted to be still in the U. S., but then got this chance to be in Poland in 2016 and 2017. Uh, and 2018 winter to just explore Europe, travel around, but also to teach data science to the youth and also improve their language skills.
And I really liked what I learned when I got a chance to visit Germany over Christmas really liked the word free education. It's something resonated with me in an amazing way. And then, yeah, I just, uh, applied to German university, got a full scholarship in one of the universities in Northern Germany.
And, uh, that's how I landed here in Germany.
Alex: You mentioned web development. So data science was not the only thing that you're interested in within this broader field of computer science, tell me about the day when you felt that machine learning and causality in particular is something that you cannot imagine yourself in the future without.
Ishansh Gupta: I mean, I did my bachelor's in computer science and, one thing I always. I was sure that it's a vast area, like computer science. There are a lot of things to do from website development to Android development. And this then comes as a big data thing with all the data science, artificial intelligence.
So I wanted to explore different things, try out different things, and then wanted to see that what suits me. So I started, through website development. Uh, then I went to Android development and then, uh, I got a chance to going to data science, absolutely love data science, working with numbers, predicting stuff, and just being ahead of the game.
It started with my fascination of just, uh, predicting something with the stocks, because I'm also into stock investments. And then also just predicting that, uh, if my team will do, good or bad. So it just started with, uh, with a hobby, then it became profession. And then I always felt that there was this missing thing in data science, especially when it comes to the explainability part.
There's always black box modeling. And in the end, it's just mere predictions and not business recommendations. So one thing also about me is I come from an entrepreneurship background. So I always like to keep my entrepreneurship hat on. So whatever may come the situation. So when I am a data scientist, I'm predicting something working in a corporate.
I'm not just thinking as a data scientist predicting, I want to make a difference. And I want to have some actionable insights and that's. One thing I always felt missing and then I came into Causality and I saw that just connecting the missing puzzle and that's really what I liked, what I learned in Causal Machine Learning because it really was able to I'll say covered the areas of explainability black box modeling, and then it was It's way more than just predictions.
It was also business recommendation and then just the capability to do what if a counterfactual scenario. So that's how I came into a causal machine learning field.
Alex: What was the first moment when you heard about Causality?
Ishansh Gupta: It started with just me or Googling. So imagine just me just Googling some of the stuff like explainability or also how can input the tribal knowledge into my.
Models, which will be still different than human aid reinforcement learning. And yeah, then, uh, while I was researching during my PhD, I got to know about causal machine learning and saw some of the videos, saw some of the talks, there were not enough, and I'm talking about around three to four years ago, even then it was not a lot of stuff about, uh, causal machine learning, then read this book, the book of why by Judah Pearl , Amazing book.
And, yeah, that's when I got into causal machine learning. And that's how I really wanted to know a bit more about that.
Alex: Your PhD program was combining theoretical work, research work with applied practical side, can you tell us a little bit more about your PhD? What was its main focus and its connection to causality and causal learning?
Ishansh Gupta: Uh, yes, sure. So, one thing I also liked about Germany is they have this PhD programs where you can work in industry and then also still connected with acadmics. So personally, I've never wanted to do PhD, but I really liked when I got to know about these kind of programs in Germany, I applied to one and yeah, just, just an idea to still do some, practical work from what you have learned in the research papers.
And it's not just researching, it's also development was super exciting for me because I'm usually one of those guys who will read something on LinkedIn or some of the research papers and really want to do it right away. Where in a conventional PhD, sometimes you just don't get a chance because you are really stuck in the literature review.
You really, really, really want to go deep into that and it's still going deep, but it still gives you the chance of, um, practically applying what you have learned. And best part is seeing right away the results in the industry that how it can affect some of the business decisions. Maybe we can end up saving a lot of money.
Maybe we can influence some other things. So that's. Uh, pretty exciting for me.
Alex: What was the main topic of interest that you focused your attention on in your, in your PhD work?
Ishansh Gupta: Yes. So, uh, I come from supply chain background. Supply chain has always been interesting for me also. Why? Because where I did my master's, uh, from Northern Germany in data engineering, a few of my, uh, batchmates, they, they came from the supply chain management, and we also used to have some of the classes.
And I really liked to know a bit more about supply chain because I saw a lot of use cases combining, machine learning there. So, I really wanted to go into supply chain management and. Right now, if you see the supply chain, it's really complex, I think more than ever, because you combine things like the geopolitical issues, or you talk about the pandemics like Corona, or you talk about the semiconductor issue, which is specifically affecting the automotive industry a lot, I see it, these are urgent problems that needs to be addressed.
But it's also an opportunity for the companies to be ahead of the game. And, uh, that's, that's where I see a big use of machine learning, especially causal machine learning, coming into play, helping the, business requirement owners, uh, make better data backed decisions, but also not just predictions, but business recommendations.
Alex: What challenges do you think causal machine learning can, can address in supply chain? The traditional solutions, statistical machine learning solutions cannot address.
Ishansh Gupta: So, I'll talk it in a way of some of the conversations I had with the management or, for example, I worked in different industries from, uh, sports industry to now working, uh, in automotive.
When you talk to these people as a data scientist, you go to them, they have a big responsibility to make big decisions. And they'll always ask you the questions like, okay, explain me the model. Talk to me about that, what exactly cost something and tell me the next step. And it seems a very obvious question and it is, but sometimes the data scientists still struggle explaining them about the accuracy scores, the root mean square.
Then you talk about the postdoc explainability methods like LIME and SHAP. And this is exactly, we start losing those stakeholders because they just don't believe in them. They don't want to play around with it and that's where the gap is. So what I felt or what I wanted to research upon is that how about we try to compare the traditional descriptive and diagnostic analytics with the predictive part.
And then the causal analytics and then see that how stakeholders feel about that. And the results we got were really satisfying. But yeah, I mean, coming to the problems addressing part, I think the explainability part, they're really explainable, the models, because of the causal discovery graphs. Then you can run these simulations, the what if simulations.
You can play around, have different simulations. So not only historical simulations, but also the future scenarios, which gives stress to the management. In the end, I always try to relate it with the trust factor because the stakeholders, they trust it more. Why also is because their knowledge is sometimes in the models.
So they actually see it as an extension of their own brain. And, and, and they really like to play around with it and they really enjoy it, and maybe the models won't be the best or the perfect models initially, but eventually they do become because the management, the trust in it, and they want to make it better.
Alex: I had a very interesting conversation some time ago with Juan Orduz from Wolt or Wolt. And he shared with me that in his practice, this is very common to build causal models in an iterative fashion. So, rather than thinking about just defining all relevant parameters in advance and then running some computations and going home, we would rather build something maybe overly simplified but something that we are pretty sure about and then iterate over this, triangulate the results of the model with real data over time and collect more expert knowledge over time and improve the structural information up to the point where the model gives us something useful. Is this iterative approach something that you could also relate to in your work?
Ishansh Gupta: I do. And I think it's a really good strategy because initially we, as people, we try to hit the home run every time, but in corporate, it's not always the best strategy.
Why? Because people you always need to have the trust from the stakeholders. You start small, you start simple. And I think then you build on to something beautiful. It's always the right strategy. Easy. But there's a. fine line between making something simple, but also you also want to give these stakeholders something extra because then they'll always say that, ah, okay, I don't need a model to tell you, let's say my worst supplier, because I already know it that this will be my worst supplier if I see the trend.
So you always need to give them some, something extra. But I think starting from something small, something simple is always a good strategy. And then you can always build onto it to something more beautiful.
Alex: You mentioned causal discovery at some point as well. We know that causal discovery might be very challenging in real world settings because of the limitations of the methods.
And we can have some, we have some theorems showing that those challenges are deeply ingrained in the I don't want to say in reality, but, uh, let's say in this thinking about causality, there are some fundamental limitations. What was your experience with those methods? And how did you deal with the challenge that we cannot have a guarantee that those methods give us reliable results.
Ishansh Gupta: Yes. So, I mean, I'll say these models are of course not, not the most perfect, but I think we as humans also are not the perfect. So, there are always human biases existing, but I see causal discovery or causal machine learning in general as a way to reduce this human biases. So my experience with these algorithms have been pretty positive.
So, I've done the study where I've interviewed different experts and tried to come up with a tribal knowledge graph of, let's say that how the external world or external factors or these news are affecting, say, so the backlogs of the suppliers, or you can change any kind of different use case if you want to predict something, why, and you want to understand that what are different factors affecting it.
So, in the end, you will have this beautiful tribal knowledge graph , which you can take as the ground truth. So what I personally wanted to test it out that, I just throw in the data and, let these algorithms run this causal discovery algorithm and then see that how close it to the ground truth, because now I already have this beautiful ground truth.
And the results I got was really satisfying which I was pleasantly surprised. I tried out different algorithms, saw some of the results, which were really close to the ground truth. But I also always feel that you need a business expert at some stage always who can testify that, okay, this is something that makes sense.
Or maybe you correct this particular edge and it will make more sense. And it can even get the business expert thinking that maybe this could make sense. And, uh, that's always the exciting part. So for me, I think I'll say that these algorithms are not perfect , but like I started, so are we.
Alex: What would be your advice to people who would like to apply causal discovery algorithms in their use cases? What would be the main lessons that you would like to share with those people?
Ishansh Gupta: I'll say just first get your hands dirty and just try them out. What always is necessary is you have to also look at what kind of relationship is between the data, whether it's linear, nonlinear.
Uh, and of course you still have to follow the conventional methods of, um, data science, data engineering in general. You have to work with the clean data. You have to do all the stuff. But yeah, getting hands dirty, try out, try out different things, try out different visualizations, and then Have always this domain expert try to validate it.
It's definitely will be one of my biggest lessons and the teachings that I'll give to the people. Try out different algorithms. I think there are some really good algorithms existing open source and then see how the results are.
Alex: For this part of our audience that is interested more in technical details, which algorithms worked well in your case?
Ishansh Gupta: I think PC algorithm is always a good one, particularly for my use case. If I have to share. So I use this algorithm called RESIT, performed really beautifully well, where I wanted to compare that how the tribal knowledge graph that I obtained from interviewing the experts, like around 25 experts. Uh, can I get something close to by using these algorithms and RESIT performed really, really well.
And what I wanted to predict is get this causal discovery graph of how different internal supplier performance KPIs, let's say backlogs or wrong deliveries or special transports, which are very general KPIs and supply chain. Uh, but they're still not straightforward that how they are connected to one another.
And, uh, the Causal discovery graph I got was really, really close to the ground truth. But that also means like some other algorithms, like I said, PC and all, it doesn't mean that they are not good or bad. It just depends solely to that particular task and that particular thing that you want to predict or if it's suitable for your kind of data.
Alex: What have you learned about interviewing the subject matter experts during this process, you said you interviewed 25 people. What would be the main lessons that you could share with the community to make this process more effortless for them?
Ishansh Gupta: So interviewing experts is I'll say one of the more necessary things and it.
It's really necessary, uh, or it can be amazing in, uh, causal machine learning models later on. So intruing them, uh, first you have to do it in a very non biased way. So maybe I can also recommend some of the methods like, uh, multi criteria decision shipmaking methods. Specifically, you can use FAHP or TOPSIS or, analytic hierarchy process.
I think they're really good, which ensures that there's some consistency in when you are intruding the experts, there are no human biases. And then you have this normal conversation with them and try to. Find this cause and effect relationships. Let's say, for example, in my use case, I wanted to know that how the external world is affecting the backlogs of my suppliers or different KPIs of my suppliers.
So, so you talk with them, you try to together build this tribal knowledge graph with them and then testify. And I think one thing also is that usually even experts won't realize initially. That they know that, that much. So that was a very interesting observation that I found and I'll sit down with them for a hour or two workshop.
And in the end we'll have this beautiful DAG graph and even they'll be like, yeah, okay, this is really, really amazing. So one thing is just keep asking those questions, motivating them to tell more things in detail and , yeah, in the end, like still get that particular graph validated and make sure that it's a DAG so that it could be utilized in causal machine learning.
Alex: We had a dinner together yesterday, and I remember you said one thing that I really, I really loved, that is related to what you're speaking about now, that constructing those graphs is a way, of course, it's a way of gathering information, valuable information, but also it's a way of In a sense, building a monument for the knowledge of people in the organization.
So even if somebody was for 30 years in the organization, maybe this person is retiring now or just changing the company their knowledge. The legacy is somehow saved in this, in this graphical structure. And I found it very beautiful.
Ishansh Gupta: So, yeah, I mean, talking about that, like it's always beautiful for me to know that, for example, I've worked in different industries, different companies, and just to know that maybe some of my work is still existing in the company and it's creating a difference is always amazing.
And trying to relate it with causal machine learning or these graphs is that I always imagine it has a chance that to utilize all these abundance of knowledge existing in these experts. And just utilizing in making a difference. Um, uh, and even when they leave a company, that knowledge somehow still stays within the company.
And I think it's really, really beautiful. And that's where even the companies in corporate, they'll have the best usage of, of, of these experts. And, um, yeah, I think it's already a kind of a plus situation and a win win situation because you as an expert will be thinking that, okay, my knowledge will always be existing within the company in these models.
I'll always. Even though I'll leave the company, I'll still be able to make a difference. And I think this gives you happiness. And also for a company, I think it's beautiful that they've invested a lot in a person, and their knowledge can be utilized and able to make these differences. So the impact is lasting.
Impact is lasting, and impact is lasting forever.
Alex: Many people in the causal community are considering today, The ways that we could leverage the power of generative AI in order to make causal learning, causal processes, causal discovery more efficient, faster or maybe just better. What are your thoughts about this intersection of those two fields?
Ishansh Gupta: Exciting to begin with. And yeah, I've tried to explore that too. So, uh, one of the challenges, uh, or, uh, I'll also say when I used to interview these expert was it takes a lot of time. To gather that meaningful relationships and the DAG, it takes a lot of time by asking them a lot of questions and, for example, in general, it will take at least two, three hours to intrude an expert.
And if you are talking about entering 25 experts, it's more than 50 hours. It's a lot of time. And sometimes in corporate, you don't find that much time. Now, the role of LLMs, how I see is that it can be definitely used to expedite the process of causal discovery. So now what I'm saying is that instead of starting from scratch, uh, with these experts where I'll have little to no knowledge about, their domain, I just go to LLMs and try to find these causal relationships, build already a causal discovery graph.
And then let the expert validate it, or he can criticize it, which is also good for me and eventually they'll build onto that because what we have also learned is that people really like to criticize. They really like to, but also they want to validate things also because they want to show their knowledge and.
I think it's a good way of, instead of starting from scratch, have something and then build something way more meaningful to that. And I want to share this experience is that, uh, the research I was working on. Where I was interviewing these experts more than 40 to 50 hours, saw these interesting papers with LLMs and, uh, Causal AI, then really wanted to utilize LLMs. I was not expecting a lot because what I wanted to do is that maybe I use LLMs to see that how the external world is, can affect some of the suppliers, but the result I got was really close to the ground truth, just after one or two prompts.
So that really got me thinking is that just after two or three prompts, you can get so amazing results, which is so closer to the ground truth. And that definitely makes sense of instead of starting things from scratch, using this as a recommendation and then build something more meaningful, meaningful to that.
Alex: So it's, it sounds to me like it's not only an efficiency booster, it's also an additional element that can motivate your experts to give you more of the knowledge, uh, more of their knowledge because they want to show that their knowledge is there and they want to share it with the world.
Ishansh Gupta: Uh, yes. So, uh, I think even they had some of the surprising reactions to what, what they learned, what the LLMs can do and the potential. But in a really good way because, they, they appreciated what these models were recommending in the causal discovery graph. But yeah, of course they wanted to add more into that because they wanted to show, show their knowledge.
They wanted to show their unique experiences, which I'm pretty sure in every industry, every expert will have some of the unique experiences, uh, despite of whatever length they have been at their company. But yeah, it was really satisfying to see that thing. And yeah, I think it builds more trust. And I think future of most industries personally, I feel, and especially in supply chain, I see is it's not just about better technology or more technology, but I think better trust.
And I think causal discovery really, really helps us bridge that trust deficit.
Alex: What are your thoughts about evaluating systems like this?
Ishansh Gupta: So evaluation, I think it's still a challenge even in a causal world, but that's where I've always felt the role of stakeholders or these business experts. becomes even bigger.
It's definitely, it does not reduce, but what happens is that we make things simpler for them by making this causal discovery visualizations where they can see all these edges, where they can see the effect it can have when they take a certain action. They can run simulations. And I think simulation capability, this counterfactual capability is something Always management wants, and there's always a causal question in the end by the management.
So if they can already run through it, that's where even though initially the model won't be perfect, it will be eventually when they start playing around with, with it. And that's where I know that there are these data evolution metrics also in the causal world, like talking about the causal discovery methods, we have the SHDs, hamming distance, number of indirect edges, but I like to read on those numbers, but not totally depend on that because eventually, like I said, the role of the stakeholders and these business experts, it's still a lot, and especially in, uh, the causal world, I think it becomes even more.
Alex: At some point in your career, He worked for a famous German football club called Werder Bremen. Do you see that causality can also be applied in the context of sports?
Ishansh Gupta: Yeah, I think it's, it, it definitely is not restricted towards any particular industry. So for example, if I want to know that maybe.
What caused a sports person injury. So doing this root cause analysis and also I could maybe play around with some of the variables. So if I want to say that, okay, X, Y, Z player, what if this person will sleep? eight hours instead of the six hours he's sleeping right now. What effect it could have on the match performance, let's say the expected goals, which is a really important KPI that, that we try to track in the sports industry, or maybe just, just in him getting injured or not, which is again, a really important KPI because in sports industry also, we always want to know what's the status of the person, whether he'll be injured or not, because that could really make or break a team. So, uh, I think the potential is immense and not only restricted towards a specific industry like sports or automotive, it can be applied to pretty much everywhere.
Alex: Is the future causal?
Ishansh Gupta: I think it's already the present. Particularly I see that it has been. Highlighted way more because of the recent events and yes, specifically relating with the supply chain, with all these issues that I mentioned earlier with the Eastern European war or, uh, these, uh, pandemic and the semiconductor issue definitely got highlighted a bit more and people really realize that there is a need of something more and there's something missing about the, these, these are black box models in the end.
So that's where I feel that a lot of companies have gone into this causal world, exploring things and even implementing it right away. So I'll say in future, it will be even more. And lately I've seen the trend in the past one year. It has definitely expedited, but I see definitely a lot of companies using it right now and in future it'll only increase.
Alex: What would be your advice for companies that are interested in applying causal methodology, causal inference, causal discovery in their work, but they don't know what should be the first or the first two, three steps to take?
Ishansh Gupta: So I see an interesting thing in the company is where it's very common for them to go this descriptive diagnostic way to predictive and then the prescriptive part.
So I always like to associate causal AI with this prescriptive part more. And in the end, like it's always about what the business wants, and that's where the causal AI comes into the play. So my advice also always is to why not move straight forward into this world of. actionable intelligence via causal intelligence, instead of just going this latter way, because eventually what most of the companies already realize is that when we do this predictive way, uh, we don't go beyond the experimental phase a lot because of, of, of these common problems about the explainability and and the stress deficit and, um, people don't trust these postdoc explainability methods. So just trying out this actionable intelligence and it definitely needs a bit of research too, because it's not that straightforward. It's something new also for the, most of the data scientists working in the industry so. So to start with a bit of research, but then trying it out and what I personally felt is when I've talked to the stakeholders, it could be also the business experts, the data scientists. One thing everybody says is that it just makes sense. And I think even if these companies are not using it right now, they will eventually because of the, of these issues.
So they don't have a choice and uh, I think, um, Causal AI is something that will definitely help them.
Alex: Many people starting with causality, experience something I like to call the fundamental fear of causality, which is asking the question, how do I know if my DAG is correct? What would you say to those people?
Ishansh Gupta: So, I mean it also comes to the evaluation metrics part. So that's where I'll potentially answer in the same line that yes, you can always throw them these numbers regarding, uh, the Hamming distance, the SHDs, number of indirect edges. You, you can do your best. And I think it's always, always required because not only you have to talk to the stakeholders, but you have to discuss it with other data scientists.
And that's where you can throw them all these metrics. But eventually it just comes down to the business experts. So what we also are doing is that a lot of people are always scared when we talk about causal AI and all, because it's just so good. So there'll be, are you replacing my job? But when they actually try it out there, they always think that it's extension of 100 percent their own knowledge.
So just involving them a bit more in this process of the evaluation is always a, always a good idea. And I think that's always a way where they will be like, okay, we are pretty sure that, okay, this edge makes sense. But okay, this thing definitely does not make sense. It's always also a good idea about interviewing multiple experts because then again, we'll have this human bias problem.
So always good to have, in true different experts. In the scenario, yes, if we don't have experts, then it's a very different scenario, but usually in corporate, of course, you will have the tribal knowledge and the experts. But in that case, definitely you can only give the data evolution metrics.
Alex: You mentioned that one of the challenges you see in broader application of causality is that data scientists themselves might not be familiar with the methods.
What would be the resources that you'd recommend data scientists interested in this topic to start with?
Ishansh Gupta: So I think it's very mental thing also, because it's something hard to believe that something like this exists. And when you talk about all these assumptions and do calculates, it's something, And most of the people, they have not learned it in their education.
So I'll always recommend them to start with this book of the book of why, which also will change their mindset. And I think specifically lately, this book, uh, causal inference and discovery in Python by you, I think it's, it's a revolutionary book. If I have to put it that way. Uh, because one thing is people of Python and, uh, about this book.
One thing I really like is also learning by doing, you see all these examples. You first learn their theory. And, uh, then you just apply right away and you see the results and I think that's where people will have more ideas. They'd be like, okay, this is something I can associate with my use case. Let me try this thing.
You talk about the connection with the LLMs, then you talk also about NLP and. When you talk about all these familiar things, the data scientists, people will know that, okay, it's not a foreign language. It's just you are learning to do something in, in an even better way and there's no harm to that.
So initially I'll definitely go the more, the mental way to remove that particular barrier in the, in the mind and then just test it out. And, uh, I think, a lot of data scientists struggled in the test part because they did not have a lot of resources. Yes, we have these open source libraries like DoWhy and all.
But, not, not a lot, but now with, with these books and some other books coming in the market right now and amazing research papers, I think people can just get their hands dirty and try out different things.
Alex: Which research papers did you find personally meaningful or important in your journey into causality?
Ishansh Gupta: I'll say the most interesting one recently I was reading from Microsoft where they combine the LLMs and, uh, the Causal AI methods. So that was particularly interesting for me. And, again, the part of what you learn just try to apply it right away. Of course, you will not always have that chance to do that because of the time or some other constraints, but particularly I like that and I just used it In my particular use case and I saw the results and they were very much In line with what Microsoft was claiming.
So one tip always will be to keep in trend with all these research papers. And, uh, I think LinkedIn has an amazing resource and a very underrated tool. So I'll see the post from all these causal ambassadors, including you or some other companies. And I'll always be thinking about some different problems in the causal world.
And, uh, That's the answer right away. So one other tip will be to just always connect with these people on LinkedIn. Ask them questions, but because that'll also get them thinking, because I always love to read this comment sections, uh, whenever someone is, uh, posting a research paper about causality and all, and they'll be like, okay, maybe you could have done this way also, or I love the way you did it.
So, I think it's a brilliant community right now on LinkedIn and, definitely one of the underrated tips, um, here.
Alex: In your career, you, you wore many hats, so you were an entrepreneur, you worked for, for a sports team, you currently work for one of the world's leading automotive companies.
What was the role of networking in your path?
Ishansh Gupta: So networking is even, I think one of the most important things. And, the more you climb the ladder in corporate, the more you realize that it's all about getting the right people in a meeting to get things done. And you might say that it's a very easy thing, but it's not.
It's the easiest thing in the world and the most difficult thing in the world. at the same time. So even in my use case, when I had to utilize these causal discovery, get it testified by the experts or interview all these expert. You need networking because it's not straightforward. You have to convince them to talk about like, uh, for hours, like two hours, three hours discussing about their tribal knowledge.
So networking is always huge. And I also like to do a different kind of networking where I'll talk to different kind of demographics, people from a very different field. Like it could be a lawyer friend or someone who is into a medical field. And try to discuss my, my problems that I'm facing in the industry and get their point of view.
And sometimes you will have the most amazing ideas like that. So networking, I think it's really beautiful. It's really necessary. And I definitely see it as prerequisite. If you want to have success also in the causal world, because you always need this experts to validate your causal models and give you suggestions.
Alex: Who would you like to thank?
Ishansh Gupta: I'll say starting, uh, my family with the support that they've given. So I come from a very interesting background of a medical background where my parents, my sister, all of them, they, they are doctors. So just to give me the support of, uh, letting me do what I wanted to do, because I've always been intrigued with the data world since I was a kid.
And like you mentioned, like as a 10 year old kid. So I'll definitely thank my parents for that. But also like, uh, the friends. And definitely I won't forget the LinkedIn community because that has really, really affected me in a positive way and really, really helped me get one of the best ideas whenever I've been stuck, I just, just, I always have LinkedIn installed in my phone. Like I always try to read out, sometimes I'll be in a club or a pub and I'll see all these interesting research paper in the causal world and, um, always happy. So just to all these, uh, random people on LinkedIn, uh, doing the good work, uh, I think these are the people that I really want to thank.
Alex: You mentioned your parents, what would be one or two things that they did that made you feel? I
Ishansh Gupta: think the first thing, of course, was allowing me to go into this wonderful field of data rather than medical, which is a very uncommon thing if you are coming from a medical background in India, more or less, you will go into this medical field.
Second was giving me this freedom to travel around and experience different cultures and different things like allowing me to go to Cairo or allowing me to go to Poland and Warsaw or Lublin. Um, Or now just allowing me to be in Germany. So I think this was a really beautiful for me and. Just even though they don't understand a lot of my field, they're always supportive.
And I always used to, even now I have these conversations, even about causal machine learning with them because I want to understand their perspective. I talk to them a lot about LLMs. So just being willing to do this active listening part by the parents and then just giving genuine recommendations and suggestions as, uh, wonderful.
Alex: You also mentioned the, the LinkedIn community. Who are the people you would recommend to follow in, in machine learning or causality?
Ishansh Gupta: So there are a lot of interesting companies that I can definitely recommend to follow. causaLens is one of them. I think that they do incredible work. If you go onto the website, they have done some really brilliant research and also in, uh, industries, they've done some impactful work. Then of course, like you, I think you have a really good community, but also your post a lot and some relevant stuff. And yeah, I mean, just, uh, I'll say just connect with a lot of people from academics also, uh, they'll, they are the ones who will always try to learn the most recent thing.
And, uh, the interesting ones also are from the industries where you'll see a different kind of use case being applied, like people from Spotify or Netflix, you see that, okay, okay, maybe I can apply this thing in my use case too. So I'll say follow different kind of people with coming from different demographics, different companies research, um, academics and industries.
Alex: You mentioned academic people in many of my conversations. There is this topic, that some people, even though some of them studied things like statistics, they learned about causality much, much later, after they've graduated, they've finished their studies. Because the programs were not talking about causality, even if causality Would seem for us, maybe a relevant topic within those programs.
Do you see a change in academia as a person who is also a supervisor now for, for younger generations of students?
Ishansh Gupta: I do. So starting with myself, like I definitely encourage my students to read a bit more about that. About causality and this whole new world, but also like I've had a chance to, work with and go to different universities, be it, uh, TU Munich or LMU or MIT and I really enjoy these conversations, with the current students, the PhDs or researching right now and Causal machine learning is a big technology in their work.
So I think I definitely see the change coming in and it's really, really exciting. And I think it will only get better.
Alex: I think that's great news. And , I'm so happy to see the new generation having this opportunity, at least in some of the universities now to learn about causality during their, during their studies, because this is also the time when they might have a little bit more space, at least some of them to go deeper into this topic if they feel interested. And then when they start their career, they're in a very different place than, than, than we were, in the beginning. You also mentioned your work in industry and in particular and the idea of, different levels of analytics, starting with descriptive and going to predictive and prescriptive what are the main insights from this perspective in the context of causality.
Ishansh Gupta: So I did wanted to compare all these three levels and then just interview the experts of what they, what they feel about different analytic tools with these kinds of technology. The insights were really, really amazing where all of the experts, they preferred a causal approach and that's where I also formed a analytic hierarchy process. Again, this multi criteria decision ship making, where we also check the consistency level of these experts, that how biased their answers are or not, and interview different multiple experts in a research way. So the answers you'll get, uh, will be pretty accurate.
For me, a few of the more interesting insights were, of course, uh, everybody preferring a causal approach, but things like explainability where experts clearly did not trusted machine learning that enough, but the really trusted causal machine learning or things like stress test or time to recover capabilities, which is super important in supply chain management.
Causal machine learning scored really heavily on that. What also I discovered is that if you combine all the different criteria on which I evaluated these technologies, it could be business, it could be explainability, time to recover, stress test, and all these criteria. Around 48 percent causal machine learning is a big winner, but you leave it around 24 percent for the descriptive part and then the rest two.
The predictive part, so there is not a much difference between the descriptive and the, the predictive part, which could make you wonder that maybe if you invest a million euros as a management into this predictive part, uh, stakeholders are not willing to believe in that. So for me, what additional benefit I'm getting rather than it just being a fancy word, AI data science.
And then I'm just not willing to trust it more explainability part. I cannot explain it more. And results clearly show, showed us this thing. So what, according to my research, I got to know is that these experts for the first time, they didn't just like these results or did, uh, gladly they did not doubted it, but they absolutely loved it.
They absolutely loved what they learned. They absolutely loved the causal machine learning, which was a game changing moment for me because these experts really wanted to play with the tool. They, uh, they still play with the tool. They run with all these simulations and come out with different Interesting insight.
So that was something amazing that I observed from, from, from my experiment.
Alex: What's next for you?
Ishansh Gupta: I'm still very intrigued with what LLMs have to offer in this world, especially combining Causal AI with this Generative AI and LLM world. It's definitely interesting. And I also feel that, uh, the future of dashboarding it's definitely having this causal AI plus LLM thing. So I'm imagining I'm in a management role. I'm the decision maker. I go to this, my phone or my application and type out, okay, tell me the next player, which will be, uh, get injured because maybe his stress level is increased off XYZ maybe someone from his family is affected, or maybe he met with an accident, or I want to know that there's a tsunami in Miami.
Tell me which supplier will be affected. And in what way, so to quantify that effect also, I think that is the future. And I think, I think this will, uh, come sooner than later and I'm really excited about that.
Alex: What question would you like to ask me?
Ishansh Gupta: How would you evaluate your causal models?
Without an, without an expert. So yes, I answered it briefly, but I want to know that how you can convince the management. Would you do some sensitivity analysis or how would you tackle this problem?
Alex: It's a great question. I think sensitivity analysis and, and partial identification and all those like advanced identification strategies that allow for some types of hidden variables in the model, these are highly underrepresented methods in industry and underrated. Of course, they also have their limitations, but they allow us to expand the possible universe of use cases where we get valid causal inferences very significantly. So that's one thing, and I think what I see in when working with my clients and what I hear also from from my guests in the podcast is that iterative approach to building a model is something that works in many cases, maybe not all cases, but many cases, which means that we might build a model, then compare the distribution from the model model.
With, uh, with the observational distribution, then if we are able to perform an intervention, we intervene maybe on like a small subsample. We generate the data from the model and the do operation and we compare it and so on. If we don't have experts, this, this process is usually longer. It just takes long, more time. But if we have expert knowledge in some kinds of documents and so on, this, this knowledge can also be retrieved. So we can consult the documents using LLMs that you mentioned. And I think one of the paths that are relatively unexplored in this intersection between causality and large language models is using retrieval augmented generation. So we all know that LLMs can hallucinate and RAG, so retrieval augmented generation, usually reduces those hallucinations significantly. And I think this is one of the paths that is also, also promising, to wrap it up.
I think, the readiness to treat this as an iterative process and applying those iterations, applying interventions and comparing this with, with the model. These are some interesting paths and definitely, partial identifiability and sensitivity analysis are, are also ways to go. Regarding interventions, there's also this very interesting avenue called optimal experimentation theory where you can, assess what actions you should take in order to minimize the uncertainty, in an optimal way. So there are papers like ABCI, active Bayesian causal inference. And I know that, some people are working now on extensions of this with less limitations.
And finally, the fourth thing is causal data fusion. So if we have some experimental data already and we have some observational data there is a good chance that we will be able To draw causal conclusions from from this data by combining it and leveraging this structural properties of the data set.
So, yeah, I would say that those four things, although Three of them are maybe not necessarily, directly about evaluation. They can help us understand how good we are doing with with the model.
Ishansh Gupta: I'll ask a follow up question on that, uh, about this the data metrics is the world we live right now is a lot of black swan events keep on happening and right now, nothing surprises. So I had an interesting conversation with a few of the people from my company and we were talking about how about if there's an alien attack, I want to know the effect of it on my supply chain. How do I do it then?
Because then even an expert won't know, um, what, what could be the effect or, um, he cannot testify or verify anything. Where does causality play a role there?
Alex: That's a great question. So I think we can take a couple of perspectives here. One of them will be that the alien attack, uh, is a new variable in the system that we just haven't included because it never happened before and we just didn't know that there's a variable like this. Now, of course, aliens might impact our supply chain directly, so they just can come and maybe destroy a production line or do something like this. And I think in this sense, it would be very difficult to model , but there might be also something different.
So maybe this alien attack is just disturbing or, uh, interfering with certain processes that are indirectly related to our supply chain. Now, if we take into account the Markov blanket of our process, so the set of variables that we are interested in. That they can impact our treatments and our outcomes.
Uh, then perhaps we can model the alien attack as an exogenous distribution. And if this is possible, our causal model can still, do , a good job because we'll just propagate the distribution through the structural causal model. And, and we should, should still get relevant outcomes in this case.
Now one of the challenges you could say here is that this requires a very rich specification of the causal model. In particular, we need to know the functions in the causal model and if we only approximate them. We can have challenges with things like positivity assumption and, and so on and so on.
So we might have certain areas of the distribution that we never observed in the past. This can be covered partially by expert knowledge, but sometimes it might be challenging. So alien attack might be challenging, at least the first one.
Ishansh Gupta: Yeah, but I still feel, uh, feel, yeah, I mean, by the explanation of your answer is that it still could be useful.
So it's better than just, just random guessing or, uh, using some conventional methods to predict something. And, uh, better utilize, uh, causal machine learning, even in this unheard circumstances or events.
Alex: Yeah, it could be definitely. I mean, it really depends where those aliens are attacking, right? So, but I think, um, you know, it's also about expectations. When we think about science, maybe there's a cultural topic here that at some point I feel like. In general, in Western societies, we started thinking about science as something that is absolute that just gives us the ultimate explanation about the world, how it works and so on.
And I think this is a very unscientific way to think about science because what we have, we have hypotheses. And then we. In Popperian the language, you would say we are just trying to falsify them. And as long as we haven't falsified them, we say like, Hey, they still hold given all the data and all the measurement tools that we have today.
So we say like, this theory has a good fit in evolutionary sense, right? That's what he would say. Causal models are not something that is like a hyperscience. They have all the same limitations that science has in a sense they are an embodiment of, of the scientific method.
And so I think that it is a disservice to the community to build expectations that causal models are like the ultimate solution to anything. And if we build them, we just can close the door and the computer will, will do everything for us. We might get close to this, to this state in certain cases, business or other contexts. But I think it's healthy to also be aware that those methods have have limitations. And now if we think about strictly predictive methods, their limitations are more severe.
Ishansh Gupta: The last question from me will be if I ask you, Do you think causal machine learning is better than machine learning or it's a better way of doing machine learning or the correct way of doing machine learning?
Alex: I don't think it's better. I think, I think a machine like predictive machine learning is amazing and you know, I'm here because of predictive machine learning because one day I just read about those programs that can learn from experience and I was like, I want to learn this. It's, it's so fascinating.
I think the challenge today that we are facing is the challenge of the community sometimes missing a clear distinction between which questions can be answered with predictive associative machine learning and which questions might require a causal model or another active model and I think this is the main challenge. And I think realizing this, that we have this challenge and working on addressing it is something that can move the entire AI community , forward much faster.
Ishansh Gupta: Thanks a lot.
Alex: You did so many things in your life worked in many different contexts in many areas.
You're also a cricket player.
Some people that are starting with new things, especially where there are complex things like machine learning or causality, they might have a feeling that there is so much to learn that this is just overwhelming. What would be your advice to those people?
Ishansh Gupta: I mean, one of the reasons why I did this PhD also was it was not just learning, reading a lot of research papers, but practically trying out different things.
And just getting your hands dirty. So I know it's a lot of theoretical work, a lot of research papers, even right now in, uh, in the causal world, there are a lot of things to read, amazing books like yours. And I think there are a few more, so I, I will say it's good to research a bit good to read, but good to watch these kind of amazing podcast, but also you should just start on applying things, start on finding different use cases. I think one of the good way also is to maybe try to whatever the conventional machine learning use case you would have done, think it off. As a causal problem and maybe see that how, uh, causal machine learning can be more beneficial, put your entrepreneurship hat on and see that what different values you can bring out, uh, using causal AI.
So I think just get, get your hands dirty, do some research, but , at some stage, just, just start executing things. And that's where you'll start learning more too, because people will ask you questions. The questions will never stop. They'll ask you that how you evaluate the model. What if. There are alien attacks or what if this and that, uh, so that's where you'll start learning more.
You'll go to the community, you'll read your book or you'll reach out people to LinkedIn and trying to understand more. So eventually you'll get really good at that. Uh, but yeah, learning by doing.
Alex: What keeps you motivated?
Ishansh Gupta: Like I mentioned earlier, I come from a very interesting background of where my parents are doctors.
My sister, she's a doctor. And I always see that in this world of data science, we also have this capability to create a difference in people's life just, just, just by doing our work. And, that is something that keeps motivating me. So if I'm working in a corporate, I know that, uh, I can end up like making some really amazing models that can result in making the company a lot of money, or maybe I can make some of the football clubs win championships, or maybe I can predict that some person will have a cancer soon. So these are the things that, uh, keep on motivating me. And, I'll say that even we are kind of a fancy doctors in the end where we can predict things as early as possible and sometimes even quicker than what the conventional doctors will do.
Alex: What was the most challenging day in your career?
Ishansh Gupta: Most challenging day? I think the challenge is something that keeps on arriving every day, especially if you live in this world of production, supply chain, football clubs, the challenge never stops. So I'll say the most challenging day for me always will be where I lose a bit of motivation. Yeah. I mean, every, every day is different. So I won't say, uh, what was my most challenging day? I'll say one of the most different, uh, days was my transformation or the change from coming from a football industry, sports industry to, to automotive industry that was challenging just because of a lack of knowledge of the domain, domain knowledge.
So that was initially a bit overwhelming. But yeah, I think, once you get into that, you start learning more, you'll start doing the proper networking. You start learning more, everything, keeps taking care of that afterwards.
Alex: What helped you get going during this time of transition?
Ishansh Gupta: So I think the right people near to me, I think they really helped me. I think I've always been. I think I've always been a good leader, but I'm also a good follower. And I think I've been lucky in life to have some amazing role models, be it my supervisors. So I think, um, especially I'll give a shout out to my supervisor in BMW.
I think I learned a lot from him also. He was my supervisor for my PhD. And these are the people who motivate you because you learn from that, that maybe they also come from a humble background, wanting to create a difference in this world, trying to solve real life problems. And it just, just motivates you a bit, bit, bit more.
And, these are the people you want to be in the end. So, yeah, these are the things that really motivate me and helped me in, uh, this difficult transition from, uh, let's say not even moving from a different industry, but moving from different countries or, uh, changing different cities because I earlier used to live in northern Germany and then moving to south. Not always the easiest decision, especially since you have a lot of friends and your family's back home. So yeah, just having the right people, um, with you.
Alex: What are the qualities of a good leader?
Ishansh Gupta: So I always feel a leader should always lead from the front, but also He should trust his people. I think one of the things that I always feel is that you end up investing a lot of time in hiring these amazing people and then you just don't either trust them enough or you don't give them the freedom to do amazing work.
So fun is just be, have that mindset of trusting them more. Give them all the freedom and then magical will flow and I think always, I think what Apple said, uh, always is stuck in the mind is best ideas always have to win. There should not be any hierarchy. So if as a leader should be able to.
positive and negative criticism but also be able to have this capability to take in that criticism. So if in a meeting, let's say a leader will say something which is not appropriate, he should have given this freedom to his, team to say that, okay, maybe in this meeting you are not acting like a CEO, which is also this Netflix mindset.
So I think these are definitely the. that a leader should have.
Alex: Where can people learn more about you and your work?
Ishansh Gupta: You can find a lot of information about me on LinkedIn. You can, you can go there, read a lot about me on, or my work on LinkedIn, but also, yeah, I mean, just approach to me, reach out to me. I'm usually pretty active on, uh, this, this LinkedIn community. And if you happen to be in Munich, feel free, like shoot me a message and I always love networking with people amazing people, diverse people, it does not always have to be about causal machine learning, but just about the problem solving mindset, or maybe some of the different experiences I had in different countries, or it could be a talk about Poland.
Yeah, I think, uh, just, just feel free to reach to me on LinkedIn.
Alex: Beautiful. Thank you so much. It was a great conversation. It was a pleasure, Ishansh.
Ishansh Gupta: Thanks a lot. Likewise. Congrats on reaching the end of this episode of the Causal Bandits podcast. Stay tuned for the next one. Bye. If you liked this episode, click the like button to help others find it.
And maybe subscribe to this channel as well. You know. Stay causal.