Data Science Leaders | Episode 14 | 38:29 | August 03, 2021
Communication in Data Science: Know the Data & Know the Business
Gaia Bellone, SVP - Head of Data Science
KeyBank
Data Science Leaders | Episode 14 | 38:29 | August 03, 2021
Gaia Bellone, SVP - Head of Data Science
KeyBank
As a data scientist, you must be able to explain complex ideas in simple ways. Knowing your data, knowing the business, and presenting the data clearly to business stakeholders is an essential part of the role.
Gaia Bellone, SVP - Head of Data Science at KeyBank, has a passion for leading and training her data science team. Her priority: ensuring that her team is successful at communicating data effectively.
In this episode, we discuss:
Welcome to another episode of the Data Science Leaders podcast. I’m your host, Dave Cole. And today our guest is Gaia Bellone. Did I pronounce that correctly, Gaia?
Perfect, Dave.
Thank you. Yes, she hails from Italy, but she currently works at KeyBank as the Head of Data Science. She also spent seven years at Chase, and prior to that, she got her PhD in statistics from Carnegie Mellon. Welcome to the Data Science Leaders podcast, Gaia.
Thank you, Dave. I’m super excited to be with you today.
Great to have you here. There are three topics that we’ll be talking about today. The first is something I know is a passion of yours, it was talking about “know your data.” So obviously data is in the name in terms of data science and it is the material that we data scientists need in order to build our models. It is a key material. And we’re going to talk about tips and tricks in terms of knowing your data, how important it is, and your philosophy with your team and building out that skill. But the second thing we’re going to talk about is where to begin with regards to data science. So I know teaching is a passion of yours and getting folks up to speed on data science. I’d love to hear from you on your approach to that.
Then last but not least, you hail from Italy, and you’ve been working in the States for a number of years. I’d love to talk about just the international differences in working in data science in Italy. And then, some of the differences and what we can learn from our Italian friends in terms of building out a data science practice. So without further ado, let’s talk about “know your data.” You have a PhD in statistics, but I know in preparation for this podcast, you have a refreshing philosophy when it comes to data science. Talk about your emphasis on knowing data. What does that mean?
Sure. First of all, when we talk about data science, we really are talking about the famous unicorn, the person that knows a lot about statistics and knows quite a bit about computing. Not as much as a machine learning engineer, but getting close. He knows a lot about the business. When you know a lot about the business, actually, what you need to know, it’s a lot about the data to describe the business and that can help a business or, not just a business, any field that you’re working with to try to solve the problems that are over there.
So whenever one of our data scientists gets a project or wants to work on a project, and if they come to me with just saying, I want to apply this or that model, I’m not even listening. I swear. I’m like, I don’t care, what are you talking about? That can be a monkey at this point, a monkey can write 10 different models and come up with an answer that makes sense. The first thing that I want from you, it’s talking to me about the data and why you think that this data would be useful to solve the problem that you are trying to solve. Talk to me about the business problem. Actually, first, talk to me about the business.
Tell me who you’re going to work with, what we are talking about, what we are trying to solve, which kind of data is really relevant for that specific field—and then we can start to talk about whatever fancy models we want to run. But there are plenty of tools now that you can run 10 models at one time, and it’s going to give you a wonderful answer.
But it can be total crap, unless you understand the data and the assumptions that your data and your problem and the field are going to suggest you, and then you can check it with your results to make sure that you’re not incurring them, the usual problem that you have when you don’t understand your data first.
Right. In terms of building a model, there are various phases of any data science project, and one of the first phases you have is EDA, right?
Yes.
Exploratory Data Analysis. And are you basically saying that, before you come talk to me about what project you’re working on, or you didn’t understand the business problem you’re trying to solve, but I want you to have done some of that EDA, right?
Yeah.
Before you come talk to me, what is a great Exploratory Data Analysis? What does that look like to you? Is it a visualization? Is it some formal documents, or is it just you have looked over the data, you just have some frame of reference before talking to you.
So I take all of these as an opportunity to train my people, to talk about what they’re doing. One of the problems that analysts or technical people have is really explaining what they’re doing. Unless you are able to explain what you’re doing, you can do a great job but nobody is ever going to use it. So the value is still zero, right?
Right.
So I train them in, for example, building decks. Visualize, build a story. Talk to me about the data as a story, start with the business problem. Talk to me about why this data is relevant, how it looks, which kind of issue they have, how you’re going to treat it. Then we can talk about the rest. This way it’s really, on one side, a learning moment. First of all, it forces you to merely learn about the problem.
Then it forces you also to become an expert in it. And then you become an expert in talking about it. Then your business partners are going to listen to you because they are going to feel that you are with them. It’s not that you’re just the crazy data scientists that run fancy tools, but have no idea about what they’re talking about. You become a strategic partner, because this is supposed to be the data science role.
I’ve spoken before to two other guests who have even moved into strategy, come from a data science role into a strategy role, because just through data, they’ve become that close to the business to really truly understand it, that they’ve actually moved into a little bit more of a strategic role. If you play it right and you really lean in to what you’re saying, Gaia, which is really understanding the business and understanding the data, there’s so much you can learn about the business that even the business folks who are not data scientists don’t even know about.
Yeah. You’re actually going to be the expert, and they’re going to come to you. For example, the way I usually structure my business, I have a team that does the data driven strategies, which means that they look into data to build a long-term strategy for the business that is big enough for involving the rest of the team to dedicate resources to solve that problem, but also involve all the other partners that need to support those specific opportunities. We are becoming strategic partners from the very beginning to find the next big thing for the business or the field that you’re working with every single time.
Yeah. That’s also another great point, which is you need not always start with some business idea. I think what you’re saying is that there should be a healthy mix of some amount of R&D, some amount of that research that your data science team is doing to come up with projects and actually propose those to folks on the business side—not always just take orders, so to speak, to build out products.
Yeah.
If you had a guess, in terms of your team in the last 12 months, what percentage of the projects were generated by the business side versus were your team came up with it on their own, what would that be?
So I would say 15% from the business and 85% from ours.
Really? Wow.
Well, the business doesn’t really know what is the best tool that they would need to solve something. So often, they come to us with problems that really are not well suited for a team like mine.
Mm-hmm.
So I send them more to teams like business analytics, right?
Yeah.
Stuff that they can give you in two weeks.
So it will be more descriptive, like a dashboard or something like that?
Yeah. They don’t need those for customer profiles. They don’t need us to know, “Oh, can you give me a picture of the sales for the last month.” You don’t need it for that. It’s not the best way to use our expertise. But another point that I always make with my team is, you don’t just look at the data. You really need to know what we are talking about. For example, I’m in banking. But I need to know about economics. I need to know about economic life cycles. I need to know, for example, now we are in a recession, what we are going to expect from the recession, so that we can start to look at the data and what the data can help us to support our business during the recession period. What is the shape of the curve? So it’s not just being a data scientist, you need to be an expert in math.
Right. Yeah. I mean, certainly if you’re in a bank, like economics and understanding macro trends and things like that.
They can tell you what you should look at in the data to actually help the business thrive. Similar thing if you are in the medical field. You have to understand what is going on in the medical field to come up with a good idea of what you should work on.
I’m curious, if I were on your team and now we’re building out a great deck for you. It sounds like you prefer somebody to actually build out a deck that proposes a project and it walks through the data, can you go one level deeper? What does that deck look like? What are some of the slides that have to be in there for you to be able to bless the project so that it moves to the next phase?
Sure. The first one is the executive summary. Tell me why we’re here, grab my attention. It’s really for this I’m saying it’s a teaching moment because of course, they could just show me a plot and stuff. I would get it. I did their job. I know what they’re talking about, but I’m not training them for myself. I’m training them to get better for themselves and be able to be independent and grow. That’s my role. So the first thing that I really want for them, it’s learning how to write the executive summary so that they can grab my attention. The first thing that they have to convince me of is that I deserve your time. You should be here and listen to me.
“So what?” Yeah.
Then as you’re saying, they need to talk to me about science of course, I know that they are super valuable, but unfortunately, when you talk to a business partner, they are pulled in so many directions. And it’s hard for them from the very beginning. Especially if you go into details, like often we do to say, why you should be here now. Why are we here? So, help your audience to be focused.
And what is a great way to grab the audience, is it ROI? Is it, “Hey, if we do this project, I think we can cut costs by this much. Or there’s a revenue opportunity here?”
You’re telling them, the business side, that’s one of the main opportunities. Talk about volume and talk about dollar volume. Talk about incremental compared to what you have done before, what is available. So numbers are common in their language and usually their performance is based on them. Really, build everything around your audience. We have plenty of brainstorming sessions internally, which is to just geek out, tell the fun facts, but we need to be prepared that when we talk outside, it needs to be around the audience.
Yeah. So this is a deck, right? That has this executive summary, but it’s a deck I assume that you’re telling your team, “Hey, if this is great, and once we’re done with version two, version three of this deck, we’re going to forward this onto the business and propose this as a project like that.” It needs to be consumable not just by me, who understands the data science and the details, who made the nitty gritty and the geeking out part of the project, but also it needs to speak to our business counterparts. So it needs to have that executive summary. What else do you think it needs to have here?
Usually in our case, talk about your customers, talk about the volumes. So, why did you come to that conclusion that they need your help? Okay. They see a conclusion. They get excited that there is value there.
Now, explain to them why that is true. Of course, with data, we don’t use gut feelings. We are scientists first. So use that to share with them why they should believe in what we’re saying. That’s the second part. And that’s not descriptive. You can use plots. So you can use tables. It depends on your audience. Try to learn who you’re going to talk to and what they’d like to hear, but make it digestible for them. Explain to them why you got to that conclusion. And then the modeling part, often they don’t care. I know we love to explain it and say how hard it was, how many iterations you went through. They do not care. It’s very sad. But at that part, unfortunately, we can not really make it the selling point. What is funny though, they do care because they have an understanding, but they don’t often care what was relevant.
Because that’s a sanity check. And the way you work with the business in the very beginning to understand which kind of data is usually relevant in their problem. You don’t work alone. You work side by side with them, especially at the very beginning, but then you need their sanity check, given that they’re still the expert. “My conclusion and the data that seems relevant, does it make sense to you?” Or we could have a problem of overfeeding or a problem that as soon as they change the data nothing works anymore.
Right. So the explainability, right? Not just in terms of the project, but explainability of the model itself is critical because in order to win over hearts and minds and get past the sanity check moment with your business counterparts you need the features that are highly correlated with your dependent variable, there needs to be a sanity check. So I’m curious, do you think that your teammates tend to shy away from types of models that are just inherently more difficult to explain a little bit more, or maybe a deep learning model that’s a little bit more black box? Do you find that that’s maybe a challenge?
No, at least in my team, they love challenges. So they’re really trying to find a better way to leverage a more advanced technique without losing the opportunity to properly discuss with the business about the results. So instead of shying away, “Oh my god, it’s going to be too hard,” they are so excited and so geeky, that they really want to make it work. On one side, they want to show that you have better performance that is worth the extra step. One thing that I often ask my team is please, when you build something like that, especially on the modeling, when you get to the modeling part, do the simplest one. And then feel free to use whatever you want, I’m fine. But we need to prove that the extra work that would take us to really go through the validation part of something more complex is worth it. Worth your time, not mine. It’s your time that could be used somewhere else instead.
That’s great advice. A lot of good best practices. It sounds to me too, I’m reading between the lines here, Gaia, but your role is not just to help out with the the geeky parts, but also to help just guide your data scientists, to make sure that they understand the business problems, to make sure that they understand the impact that they’re having. If you can explain a fairly complex thing, it’s actually a very hard thing to do. It’s a skill that you have to learn. And I imagine, learning from you and eventually when you are able to routinely explain very difficult things and to simplify them, I have always found that I learned a lot from that process. I learned sometimes, “Gosh, I really don’t understand this as well as I thought I did, I’m going to have to research this.” That’s a really, really important skill.
Data science has moved beyond…it’s become more and more accepted. It’s become this field and this necessary role that you need to have within your organization, but you need to make sure that it’s always driving value that is relatable to the business side. If you cannot do that, it could run the risk of the data science teams shrinking.
Maybe we can switch gears. Clearly your passion for training and providing guidance as not just day-to-day manager, but also from a data science perspective is coming through. I’m just hearing it in your answers. What advice do you have for other data science leaders out there, who are trying to build great data scientists and trying to get some folks started in the world of data science as a way of building out their team, instead of just hiring people straight out of college who have a degree in data science or related fields?
So training is very important and it’s twofold. On one side, the leadership or whoever built the training needs to be a practitioner, so that the training is going to be effective. On the other side, the leadership needs to really support the training moment. We cannot expect that the training is going to be an afterthought. Unless you empower your people to be trained, they’re not going to do it, they’re going to always say, “Well, I need to work versus I need to train.” You need to give them the opportunity of actually taking training. One thing that we have done in my team, we actually put the training as part of the performance review, so that people would feel not compelled, but it was like, “Okay, I’m spending time on something that matters for my career and is included in my career path.”
Often companies give training, but they don’t really make it part of your career path. You take it, you don’t take it, your issue, your problem. If you want your team to grow and to keep up with the skills that you have to give them the chance to do it, you have to give them the chance to make that part also part of their career path. When they start from the very beginning, if they’re coming out of school, one thing that I often do, and actually I did it myself when I joined JP Morgan Chase, or right out of graduate school where my thesis actually was on DNA sequencing. I had an undergraduate degree in economics, so it’s not that bad, like finance and banking and in quantitative finance, so it’s not that everything was new, but it was five years before in a different country. So I was like, “Okay, well, how do I learn this? And how do I learn this fast enough? How do I understand that? What does it mean to be a data scientist—or at the time we were called modelers—in pretty much the biggest bank in the United States and in home lending?”
What I did over there was I asked for a meeting with everybody on the team. I started just with the teammates and I asked them to share with me a deck or a full analysis that they already have done before. So completely done work. They didn’t need to do anything extra, just walk me through. So I would have an idea which kind of work the whole team was doing. And also what is the expectation in delivering that work?
Then the second piece, it was a really little bit more of open discussion around the project that they were working at the moment, but still not trying to take time off of their day, not asking to prepare anything new. It was really for me to have a very quick overview of what everybody was doing, how that needs to be presented, what kind of work our team was going to be asked to do. It was also a very, very good way to meet with everybody.
That’s a great piece of advice, right? So if you’re joining a data science team as a practitioner for the first time, tell your manager, “Hey, I’d love to sit down with each data scientist and just get a little bit of a sense of a completed a project they’ve worked on, or maybe what they’re working on right now, just to get a sense of the expectation, I imagine. Models they are going to learn, the business challenges we’re solving, get an idea of what data, how do we get models in production?” You’re going to learn a lot by actually rolling up your sleeves and talking to each individual or team and what project they’ve done. So that’s a great piece of advice.
And you also get all the decks afterwards. So you have all the time to study the decks and take your time, because anyway it’s hard in the very beginning, it’s so much new information that you can digest and then go back and maybe ask more questions and the same kind of tactics. You can just use it to expand your network and learn more about what happens outside your team. That can be useful for being a better contributor to the whole business, the more you know, the better it is, of course. And on the other side it can open your doors.
We all know that people spend a lot of time at their job. Most people are proud of what they do, and they want to show it off. It’s also a great way to make connections, as you mentioned, network, and get to know your teammates. Another piece of advice you were saying is, advocate for your team to carve out time for training. What time do you allocate? And you mentioned that you sort of, your OKR, or you have it as part of your performance review. What is a reasonable amount of time for a data scientist? What percentage of their time should be focused on training?
It depends a lot, right? We are going through, for example, the cloud migration. We know that the team needs more time for training. We look at the book of work, we identify the training that the teams should be doing. And then we ask the team to take one course. This is how we did it also for other training, for example, when we moved to open source.
Take one class, and then let’s discuss how much time you thought you had to spend to really digest it. And then we can create a path and say, okay, for learning, for going through this full track, that they think it’s mandatory for you to really have a solid base for Python coding. I want for you to be able to do it in four months, then this is the amount of time that they think you should be able to spend. But it’s really done with the person, it is not done by us not thinking about how long it takes for a person to really digest that material.
It’s there, they have to do it. If you pretty much make them tell you how long it’s going to take, they would really dedicate the time and they will feel they were part of the decision. So they are going to actually do it. If you just enforce it, it could become a much, much harder conversation that they would fight against. Mostly because they also have a day job.
And first of all, we are trying to make the training part of the day job with this work. But at the same time for them, it’s always a negotiation. So they finish their day job. So they go to the training. Helping them to be part of the discussion and the decision on how much it would take has them commit better and find the right balance.
Bringing them into the process, sitting down with them and saying, “Hey, what’s on your plate? Here’s what I would like for you years from now. I’d like you to be at this level of being a Python developer, and this is what I think it’s going to take to get there. How much time do you think you need? Is 12 months reasonable or maybe they can do it in six months?” That’s always great advice for almost anything. The second thing I heard, too, is that you can’t always say, 10% of your time or 20% of your time should be dedicated to training. There’s certain events that might kick that off. So a move from maybe proprietary tools that your team may have been using to open source might be an event. Another event is like, “Hey, we’re moving, moving to the cloud. We’re moving from an on-prem type environment and in the cloud.” There’s a lot of services in the cloud that the team could be an expert in. I feel like there’s a whole other podcast episode on moving into the cloud as a data science.
So many stories for you, Dave, over there. So many stories.
Is there anything you could share? I feel like you need to come back on because that’s a fascinating topic. Are there any watch items for you, or maybe things that are eye opening as you’ve gone through this process?
It’s not really an eye opening for me, but I discovered that it’s an eye opening for a lot of people inside corporations. The cloud costs money. It’s a very different way of managing cost. You need to start to have the people who usually are just users of compute resources to start to think in a completely different way. Now, the problem with compute is not just capacity, which a lot of people already don’t think about. They just run their processes, never thinking about what’s going on now. Now you don’t have any more problem of capacity. Now you have a problem of cost. That’s a much harder problem to deal with because then finance is going to breathe down your neck.
Yeah.
So it’s an opportunity though, for the analytics community and teams to grow internally, because they start to have a broader view of the way they operate. Instead of just focusing on solving the problem, they start to see actually how many things that need to be included and considered whenever they are working on a problem. One of these becomes what is the right tool that they should use, how much it should cost. And then there is a huge discussion, let’s call it a Pandora’s box that should be opened at some point…who should pay for it?
Yeah, oh gosh.
How the new cost models should really look, like it should be just a look back and just making sure that people are considering costs or you want to put those costs on the budget of analytics community, or you want for these costs to be put on the budget of the the final user of the output, which is the business. I heard many different stories around there. Nobody really knows.
I think clearly the business should pay for it. That’s my two cents here, at the Data Science Leaders podcast, that’s the obvious answer. But, when it comes to moving to the cloud, my perspective is that the costs are clearer. It should be more clear, right? Because time is money, in the cloud. But it’s not like the tools that you were using in the past, whether it was on your laptop or whatever, it didn’t cost anything. It was just that they were not as easy to quantify. You might’ve bought some software five years ago that you’re still using today. Maybe you’re paying 10% or whatever per year, but now it’s just in your face and it can certainly create challenges and make that ROI conversation that we started the episode with just that much more important if you’re cutting costs. Then you have to compare it against the time it takes to train your model, or the time it takes to do this project. Obviously you should be saving more than it costs to build a model, but yeah. So I think those conversations are now starting to be had, thanks to the cloud, and it’s great that you’re going through that.
I did want to touch on you growing up in Italy, you worked there for a period of time and then you came over here. If you’re a data science leader and you have somebody who’s come from a foreign land and you’re living in the US or…this podcast is not relegated to just folks in the US…but how have you seen the differences from a culture perspective?
So the hardest difference, especially if you’re coming from Italy, it’s vacation time. We have so many days off in Italy, you get up to six weeks, but then we have a lot of holidays. You have a lot of breaks, which is really not common in the US. Actually, it was tragic for me also when I came over to graduate school, that you start them in August, and then you don’t breathe until December. And then you give me a month off. I’m like, I don’t care. I’m dead already. Which is really I think a cultural difference from that point of view in banking at least, we have not a bad number of weeks off. There are definitely other industries that are much, much stricter, but I miss it.
Yeah. I love vacation with the best of them. You need to recharge. If you’re just working constantly, I think you lose the freshness. You lose that fresh perspective.
Yeah.
Some of my best work was mentally sitting on a beach somewhere and thinking, “Oh man, I should probably do that when I get back.” I try not to think too much about work when I’m on vacation, but sometimes these things just pop in, right?
My best ideas came to me while I was walking my dog.
That’s something that folks should look for too. As people who are really passionate and live and breathe this stuff, yes, you want time for yourself and with your family and walking your dog, but you also want to hire data scientists who really are passionate about what they do, and they’re thinking about these problems morning, noon, and night. Those are the types of people that are going to be key to your business. But is there anything else to be said between Italy and the US?
We do have differences, especially about opportunities that you have in this country and definitely the data science community and the use of the data science side, business or corporation or industries in general. It’s way more developed over here than in Italy. Fun fact, when I talked to my advisor when I was in a college in Italy, and I was joking, like, “Okay, what do you want to do next? Do you want help to find a job?” And I was joking. I worked and studied at the same time. So I was tired of it. I was like, “Can I be paid to study?” And he was like, “Yeah, if you go to the US, that’s the only place that you can do that through a PhD, but it’s also worth it because of course a PhD almost kills you. Like, it’s not a walk in the park.” I still wonder why I did it and how I finished it. That’s another big question. And nobody knows, but really the only country, based on my experience, but the US really values what you can learn. And it really, at this point, probably is the place that values the most something like data science.
Yeah.
And the reason why you picked it when I picked it, it’s just because I didn’t want to pick any field. I was like, I want a very good bag of tools so that then I can choose whatever I want, because I already had my undergrad degree thesis, in Italy you need a thesis for that as well. It was in option pricing where I got super excited about stochastic calculus and stochastic processes. And I felt, “Oh my god, I don’t think I know enough. I don’t know enough that I can solve or even find a completely new problem to solve by myself.” So that’s one of the reasons why I went into statistics, but I don’t think I would have had the same kind of chance if I was in Italy. Like I could have gone to the PhD in Italy, but the PhD in Italy is very focused on theoretical and the PhD over here, you can be theoretical, but also over here, I had the opportunity to work from the get-go with actual clients and projects. That really makes you excited about what you can solve with it.
And then you’ll really have the opportunity to enter the work environment and keep doing what you love, which in Italy most likely would have been much harder to do using all the geeky stuff that I was learning on a day-to-day basis to solve real life problems, which is what scientists really liked to do.
If you’re out there and you’re wondering what PhD to get, or what advanced degree or any degree to get, I think you’re right. I never thought about it that way, but data science and by extension statistics, is so universal, it really can be applied to almost any industry. You’ve shown that. I mean, you started in DNA sequencing and then you magically transitioned over into banking, right?
Mm-hmm.
And seamlessly, and clearly it worked out quite well. So with that, thank you so much Gaia for joining me on the Data Science Leaders podcast, I had an absolute blast. I hope you enjoyed it as well.
It was a lot of fun. I always love to be able to share my experience. Hopefully it’s going to be helpful to somebody. And of course, if there is anybody that has any extra questions you can ask Dave and Dave can let me know.
Yeah. I assume people can reach out to your LinkedIn maybe and get you that way. Or if you’re on social media. Yeah, cool.
Yeah, totally. Through LinkedIn, they can find me. I’m pretty rare. My first name, last name. It’s a pretty unique combination.
Well, thank you very much, Gaia. Arrivederci!
Ciao!
29:22 | Episode 16 | August 17, 2021
40:04 | Episode 15 | August 10, 2021
26:54 | Episode 13 | July 27, 2021
Use another app? Just search for Data Science Leaders to subscribe.
Data Science Leaders is a podcast for data science teams that are pushing the limits of what machine learning models can do at the world’s most impactful companies.
In each episode, host Dave Cole interviews a leader in data science. We’ll discuss how to build and enable data science teams, create scalable processes, collaborate cross-functionality, communicate with business stakeholders, and more.
Our conversations will be full of real stories, breakthrough strategies, and critical insights—all data points to build your own model for enterprise data science success.
Host, Data Science Leaders
