Ep. 279: Tineke Distelmans - From Data to Decisions: AI in Business

(upbeat music) - Welcome to Count Meian. I'm your host Adam Larson. And today we're exploring artificial intelligence and machine learning with our special guest, Tanika Distamans, an assistant professor at Vrya University dot Amsterdam. Tanika will break down the differences between these often discussed terms and show how they impact our daily lives. We'll also delve into a case study she developed, which used machine learning to predict patient satisfaction at a hospital setting, from data quality to ethical considerations, Tanika shares invaluable insights. So join us as we uncover practical tips, discuss the challenges of AI implementation and explore the evolving world of generative AI, all aimed at making your jobs easier and more efficient. Let's get started. (upbeat music) Well Tanika, welcome to the Count Meian podcast. I'm so excited to have you here. And we're gonna be talking a lot about artificial intelligence and machine learning and some case studies that you've been a part of. And so maybe we can start off just at the top level, you know, artificial intelligence, machine learning, maybe we can talk about the differences and the, and because these aren't new concepts, everybody's hearing these terms, but we can break down the difference between the two to start off. - First of all, thank you for the invite. I'm very happy to be here. But first indeed, maybe let's say it's clear what we're talking about today. Now if we're talking about artificial intelligence that's actually quite a broad field because that refers to the development of computer systems that traditionally would have required human intelligence to perform tasks such as, for example, learning, problem solving, language understanding, all of these things. So that's very broad field where we refer to as artificial intelligence. Now if we're talking about machine learning then we're talking about a subfield of artificial intelligence. And when we're talking about machine learning, we refer to the use of statistical programs or algorithms that enable computers to learn from existing data without being explicitly pre-programmed. So without getting very explicit instructions on what to do, machine learning enables the computer to learn from its experience. - Okay, so it sounds like we encounter this every day in our lives. Are there some examples you can give that you can say you're not even realizing that you're encountering either machine learning or AI? - Yeah, that's actually true. I think nowadays we can see that it's everywhere around us in every little corner. If I'm taking my smartphone, I have my virtual assistant Siri who can understand and respond to my voice commands. If I go to my mailbox, I have an algorithm behind my spam filter determining what is a spam mail and what is not. If I go to my social media, it's an algorithm that is determining what kind of contents I see on my feet based on previous interactions that I had so that I get very personalized contents. But also if I go to my Spotify, I wanna listen to some music. I get very personal recommendations based on what I have been previously listening and the same goes for Netflix. There's an entire algorithm behind it determining what are recommendations based on my personal taste. And even my supermarket is sending me personal recommendations, let's say, or sales that might be interesting for me based on my previous buying behavior. So it's in every little corner in our daily lives, I would say. - Now, there's one thing that you hear a lot too, especially when talking about things like Netflix algorithms and so like that, that they're learning and they improve their performance. Now, are they learning in our traditional sense? Because when you think about learning, you think, oh, humans learn things, but can the machines learn as well? - The machines learn indeed as well. And this happens through a kind of iterative learning process. So what is actually happening is that we feed algorithms with data, we provide them the data. And then the algorithm will do the job. It will optimize and fine tune its parameters to make sure that it will better be able to make a prediction or to make a decision or to recognize certain patterns. And so the more data we feed these algorithms, the more opportunities these algorithms have to learn from and the better they will become and the better generalizable they will be. - So when you and our chatting before, you mentioned some about supervised and unsupervised learning algorithms. Can you explain the difference between those and what those are? - Yeah, so within machine learning, they're basically through main streams, let's say we have unsupervised machine learning, supervised machine learning. Now when we're talking about supervised machine learning, then what we actually need during the training process is we need a labeled training data set. So in a first place, we need to tell the algorithm the inputs, but also the outputs. And by providing the algorithm, the correct outputs associated with the inputs, then the algorithm can learn and to make the associations itself. So for example, if we want to detect fraudulent cases, then in the first place, we need to provide the algorithm with some cases where there was fraud or fraudulent transactions and no fraudulent transactions and we need to tell the algorithm, this is a transaction that was fraudulent, this was a transaction that was not fraudulent and so on. And then through the learning process, the algorithm will learn that association and will be able to predict it itself. Now when we're talking about unsupervised machine learning, there we don't need to label training data. We just provide the algorithm with the data and then it's up to the algorithm to find the right structure to recognize certain patterns within that data set. - Now is one better than the other, which just depends on the application. - It depends on the application and it depends on the type of problem you're trying to solve, let's say, that you decide to either go for a supervised or an unsupervised approach. - Gotcha, okay. So when we first start talking, I mentioned a case study. Maybe you can just give us an overview of the case study. Obviously, maybe everybody, it probably isn't gonna read the case study, but maybe we can give an overview and kind of give an understanding of what you research there. - Yeah, so we're talking about a case study that I developed together with some of my former colleagues at the Atlantic Business School. And so the case study is about predicting patient satisfaction in a hospital setting. Now, long story short, the CFO wants to challenge the performance of the hospital. They know that patient centricity is key and everything, but they want to challenge their performance, but also very important, they want to optimize their budget allocation. And so in order to do that, we developed a machine learning algorithm, let's say, that predicts whether in the end, the patient will be satisfied or not satisfied or about the hospital. - Wow. - Wow. - Well, did it work? I guess that's jumping to the end of the case study. (laughs) - It did work and it gave actually some very nice insights also for the budget allocation, because it's one thing to train the model, let's say, and therefore we used, let's say, survey data that the hospital was electing from their patients. So they started with putting iPads in the rooms, and then patients were asked like all kinds of questions related to different types of aspects in the hospital about the room, but also the security within the hospital, the food and the beverages that they got, the nurses, the doctors. So every little aspect was questions. And so we used that then to train the model and to predict whether in the end, the patient was overall satisfied, yes, or not. And so once we got that model, once we got that algorithm, which was performing really well, we looked at feature importance. Now, what do we mean with feature importance? So we had all these input questions that we were looking at, or that we were using to predict whether a patient is satisfied or not. And if we look at feature importance, then we look at what is now, or what are the most important features for the algorithm to make that prediction of whether the patient will be satisfied, yes or no. And if you know which features or which aspects within your hospital are driving patient satisfaction, are making that your patient will be satisfied in the end, then of course, you can adapt your budget allocation in line with that. For example, if it would come out that the security in the hospital is one of the most important drivers for patient satisfaction, then you know that security within your hospital should be at all times at 100%. - Wow, so you're able to take the data and make actionable insights. Now, let's say someone is listening to this and they're like, oh my gosh, how do I, can I do this in my own organization? Like, what tips would you give them to try to say, hey, I wanna do my own study within my organization to make better decisions? - Well, I think, I mean, if you wanna implement it to yourself, I would always recommend to have like a very good understanding of what these techniques are doing yourself. getting very familiar with artificial intelligence. with machine learning. And I think nowadays you don't even need to have the coding skills to be able to do so. There are plenty of websites online where you can just play around and train your own model without the need to code because that's all done for you in the background. But like that, you get familiar with like the training process and everything. And I mean, if you would like to learn to code, let's say there are plenty of opportunities of online courses that you can you can follow for example by Coursera or by Data Camp. And then it's a matter of just translating this into your own into your own setting into your own profession and and how it can help you over there. But also nowadays they're online. You find a lot of things for every industry. I have a lot of blocks. And for example, the website towards data science or the platform where a lot of articles appear but also very industry specific applications that appear on there. But I think also nowadays like organizations that are overseeing certain industries, they're all concerned with this matter and they're all publishing reports about it like how machine learning can be used in that particular industry, what are the benefits, but also what are the challenges, what are the difficulties maybe, what are the risks of implementing these things because these things are also very important to be aware about. But I think there are plenty of opportunities. Let's say I'm planning resources nowadays if people want to familiarize themselves with the with the techniques. Definitely. Well, with any technology, there's nothing's going to be a perfect solution. So maybe could you talk about some of those challenges and limitations you could consider when trying to implement or trying to do your own type of study? Yeah, I think there are of course multiple challenges and limitations and trade-offs that you need to make. But I think a very first important thing to realize is the thing that you need a lot of data. If you want to train a good machine learning model, you need a lot of data. I explained it before the more data you feed the algorithm, the better your algorithm will become. But not only data quantity matters, it's also data quality that matters. Because there we have this famous principle, what we call garbage in garbage out. If you feed your algorithm with very low quality data, then you cannot expect your algorithm to perform well. And so that's the first important thing is that it's not just these big volumes or big chunks of data that you need to arrive at a good algorithm, but at the same time your data should also be of sufficient quality, let's say. And I think the second important challenge is also with the trade-off, let's say that you need to make on how complex do I want my algorithm to become. I mean, it's quite impressive what is possible nowadays and how complex, if you look at neural networks and the patterns that they can find within the data, it's super complex. It's way to complex for the human brain even to see these relationships between variables and so on. And it's super impressive, but at the same time you lose some interpretability, let's say, of your algorithm. Because even the people that are coding, let's say, or that are training these algorithms do not really have a very clear understanding anymore of how the algorithm is making that decision or is making that prediction. And I think especially when you're using these algorithms, let's say, to drive your decision-making, for example, within your own organization, I think it's at least important to have some feeling about how the algorithm, let's say, is making its decisions. It's like we still need the human intelligence side of artificial intelligence. You need both, especially when you're making strategic decisions. There's these science fiction novels out there where the whole society is run by massive artificial intelligence. And I don't think we want to go that way. We still want to have the human side of things. And so when you're looking at artificial intelligence machine learning, some of the things that come up are things like biases or being ethical and responsible with the data, what are some ways to avoid getting into some of those holes that we talk about? Yeah, I think that's indeed super important to make sure that we use these type of things in an ethical and in a responsible way. But I think there are just some, if you would say, okay, let's implement this within my organization. And let's use this to drive our decision-making and to help us, well, great. But I think there are just a few important principles that you always need to adhere to, let's say. And the first one I would say is transparency. And this goes back to my previous point, let's say, of this model interpretability and complexity. I think whenever you're using these type of algorithms for making decisions, I think at least the stakeholders, let's say, of the algorithm should have least an understanding and get how and why certain decisions are made by the algorithm. Suppose that in the banking industry, let's say, they will use an algorithm to determine whether a customer is credit worthy or not. And the algorithm at a certain point determines that a particular customer is not credit worthy. Then I guess it's very important to have at least an understanding of why the algorithm is arriving at this decision. So this transparency around the model is, I think, very important. Now besides transparency, I would also say that fairness is super important. And this actually also goes back to point that I previously made about the garbage in garbage out principle. If you use very poor quality data, then you will get poor results as well. And that holds as well for biases. If there are biases within your training data, then you will get a bias in your model and in your algorithm as well. And I think that's also something that we want to avoid at all times that we create biases in our algorithm against a certain gender, against a certain race, against certain age categories. So therefore it's super important to really look at the quality of your input data and make sure that there are no biases in there, that at least in that way we can guarantee that we can create a fair algorithm. And I think a third important thing to make sure that we use it in a responsible way, let's say, is also data protection and privacy. Because we're often using very sensitive information in this type of algorithm. So making sure that the data is very properly protected, that the data is anonymized, making sure that nothing can leak, that there are no cyber threats or anything. I think that's also super important if we want to use and implement these models in our decision making. It sounds like if someone is looking to implement some sort of machine learning or AI within their organization, they have a lot of prep work to do based on what you're saying. Yeah, that's true. But I would also say, I mean, if I go back to the case study that we develop, most of the work is actually in appropriation. Also, if you just purely look at like training the model and everything, most of the work goes into the data preparation, the actual training part and developing part of your algorithm does not take that long. It's all the preparation part that takes up most time. And I think it's exactly the same whenever you start implementing that within your organization. There are a lot of things that you need to be, or that you need to think about, let's say, that you need to take into account. And indeed, I would say that the prep work is more than actually implementing the algorithm. Oh yeah, because the machines can move a lot faster once we give them the data. They're just itching to have it. Yeah, that's true. So when thinking about machine learning and AI, you know, are there things that within our daily jobs that we can say, hey, this will help improve my performance in doing things. And I know that there's lots of technology every every two minutes. You there's a dot AI or dot IO new website popping up with some new feature, you know, are the things that you've seen that work well really well. Actually, yeah, and I think it's it's quite impressive, like particularly the very recent years, like generative AI, how it has evolved, but mainly how quickly this has evolved. But it's actually that, as you said, there are quite some tools. I mean, I I almost have my virtual assistant on my laptop that can summarize email conversations for me that can then generate or write an email for me or do at least a suggestion of a reply that that I can give. It can help me to schedule meetings, but also just in generally for for writing. I think there are a lot of tools based on AI where it can really work, let's say, and help you. Yeah, for spelling checks and grammar checks and everything when it's a very important text that you're writing, but not just text, also PowerPoint slides and visualizations. And I think it's just these these small things here and there maybe, but if you would add that up that it would make your life a little bit easier, let's say, and and your job a little bit more fun because these are often like the tinier it does, but that sometimes can take up a little bit more time than expected, which can be a little bit annoying from time to time and there, I think it can really support you and help you in being a little bit more efficient, actually. Well, I think that's some great recommendations and if you're not using things, please get out there and try these new tools out, a lot of them are free at least to start with. And you can really, you know, you can really do a lot and it can help make things easier for you. Yeah, exactly. And it's just very fun to play around with it and even though you're familiar already and you have the coding skills, I even from time to time also just check out some websites and I then bump into like, hey, here you can train your own algorithm and I'm just playing around with it. And then often I would later be realizing, okay, I've been just playing around with it for like 30 minutes, but it's just super fun to do and yeah, it makes yourself so familiar with how it works, but it's also quite impressive to see, let's say nowadays, what is possible and how it can support us. Well, Teneke, thank you so much for coming on the podcast. This has been a great conversation and I encourage everybody to check out Teneke on LinkedIn and connect with her and we just, thanks so much for coming on. Thank you for the invitation. This has been Count Me In, IMA's podcast, providing you with the latest perspectives of thought leaders from the Accounting and Finance profession. If you like what you heard and you'd like to be counted in for more relevant accounting and finance education, visit IMA's website at www. IMAnet.org.