LPRC - Episode 11 – Artificial Intelligence & Machine Learning in Retail AP

Episode Date: August 9, 2018

The post Episode 11 – Artificial Intelligence & Machine Learning in Retail AP appeared first on Loss Prevention Research Council.

Hi everyone, welcome back to Crime Science. In this podcast, we aim to explore the science of crime prevention and the practical application of this science for loss prevention and asset protection practitioners as well as other professionals. Co-host Dr. Reed Hayes of the Loss Prevention Research Council and Tom Meehan of ControlTech discuss a wide range of topics with industry experts, thought leaders, solution providers, and many more. In today's episode, Dr. Reed Hayes and Tom Meehan will discuss artificial intelligence

and machine learning in retail AP with David Spikes and Daniel Downs of Atbers Retail. We'd like to thank Bosch for making this episode possible. Video analytics is an important enabler of any retail loss prevention strategy that improves security and delivers clear business advantages. With the intelligence built into Bosch IP cameras, you can use video data for far more than security alone. This includes identifying patterns in customer activity and recognizing retail hotspots to improve sales, as well as optimizing store layout to enhance the customer experience. Learn more about adding intelligence to your system in zones one through four of the LPRC zones of influence by visiting Bosch online at boschsecurity.com. All right,

well, welcome everybody to another episode of Crime Science. Today what we're going to do is talk a lot about what we call Zone 5. And as you know, we are all looking at how we define and handle our issues in three dimensions. And so we talk about Zone 1 as whatever we're protecting, that asset itself. We talk about zone one is whatever we're protecting, that asset itself. It could be a person. It could be an item. It could be cash, a cash register, a safe.

It could be merchandise on the sales floor. Zone two is that immediate area and what can we do there to protect, influence behavior, to understand better what's going on. Zone three is the entire interior space of our store or distribution center, warehouse, office environment. Zone four is the parking lot from the building wall to the edge of the parking lot. What can we do there, again, to understand the dynamics and better address ways to keep people safe and run offenders off? And again, and then zone five, of course, is beyond the parking lot in the cyber environment and into the built space, the social environment. So today, we're going to talk a little bit about, you know, fraud and other threats to our overall environment. But what can we learn out there in zone five, and then talk a little bit about a new book that's out that we're

excited about, that kind of takes a look at these things and helps all of us think more quantitatively. How do we change things into numbers and make sense of the world in that way and then convert it back? So what I'd like to do is I'm going to go to Tom Meehan, my partner in crime here from Control Tech on crime science and crime. Tom, if you would go ahead and introduce our special guest today and we'll go from there. Well, thanks, Reid. Really excited to be here to talk about data and analytics and the zones of influence. And especially excited to have our special guests, David Seitz and Daniel Downs. David, Daniel, welcome to the podcast. Thanks, Tom.

And what we're going to do is we're going to go through the zones of influence and really how data applies to retail risk and asset protection. And our guests are experts in the field, and we're going to have a lot of specific questions, but we're just going to have a general conversation like we always do. David, why don't you tell the listeners a little about yourself? Yeah, I'm David Spites. I run data sciences for APRIS. And most of the listeners are probably familiar with APRIS retail, but I actually have responsibility for the whole of APrys, which includes Aprys Health, Aprys Safety, and Aprys Retail. And then our team is probably a team of about 25, and we're focused on all three business units, but probably about 70% of our time and effort is dedicated to retail.

My background, I have a PhD in biostatistics, and I jumped into business a long time ago and sort of deviated from the biostatistics part. But, you know, my background is applying data and analytics and models, predictive models, artificial intelligence to solve real business problems. Great. Thank you. Daniel, why don't you give the listeners just a brief overview of who you are and what you do at Aperis. Sure. So my name is Daniel Bounds. I have worked for Aperis going on seven years. After my PhD in criminology, I got into LP and LP analytics back in the day with Reed. This is back when we had a little office up in the Sears building on the second floor. So I was a research coordinator for LPRC for a couple of years and then relocated out to California. And with Aprys, I do predictive modeling, analytics, and been doing that for the last seven years, working on retail and safety initiatives.

Been doing that for the last seven years, working on retail and safety initiatives. So I'll turn this over really to both David and Daniel. But I think when you talk about Aperis Retail, I'm obviously very familiar with it and excited to have you on. But could you give the listeners kind of an overview of what Aperis does and some of the products that are related to retail and actually outside the retail sector so folks can understand a little bit about what your organization does. Sure, I'll handle that one. So, APRIS retail is really the combination of the retail equation and CIS Republic, and we joined together about two years ago as a single organization under the umbrella of APRIS, which is a Louis actually, we track 80% of the opioids in the United States, and we monitor those and report them back to 42 states use us exclusively to provide that reporting back to doctors within the states. And we have a layer of analytics on top of

that. So we're trying to detect people are going to overdose from overdose from opioids you know while we're trying to fight crime and retail so we're doing quite a lot of different things and then in our safety business Aprys actually has about about the same amount of the jail bookings in the United States about 80% of the jail bookings we have 48 states that use our service in one way or another and we monitor people going in and out of jail so we can notify victims, report that data back up to criminal justice investigators like the FBI, the CIA. They're using our data to go and investigate people's arrest histories. And then we layer on top of that as well, crime analytics, recidivism analytics. We look for,

you know, likelihood someone's going to commit a specific kind of crime. And a lot of that techniques, a lot of those things we apply is exactly what we're applying in retail. And so our retail business has a number of main products there. The retail equations main product was a return fraud detection product called Verify. And that product monitors returns in real time and applies predictive modeling, machine learning type techniques so that we can detect in real time

when someone's either committing fraud or abuse or is a general outlier to the business. And then we shut that person down or we warn them, you know, they can't make returns anymore. CIS Republic, who we joined forces with two years ago, and I say we because I came from the retail equation, but now I represent all the companies. When we all joined forces, CIS Republic brought to the table a very popular exception-based reporting system called Secure. And that product we've now, my team's gotten heavily involved in. We've started to apply as much analytics to it as we can.

And we began going down the path of integrating in artificial intelligence-based, really machine learning-based exception reporting or employee fraud or abuse detection. And so I think that a lot of the questions in this conversation may revolve around that really that new step, which is to take analytics to or take machine learning and modeling to detecting employee fraud. Great. That's really helpful. Actually, I learned a lot just in that brief overview.

Would it be fair to say that would you classify yourselves as a data company? I would like to call ourselves a data and analytics company, but we are definitely a data, like as a whole, as Aprys. I mean, we warehouse data for 42 states for opioids and other 48 states for jail data. And for retail, we're probably a little bit of both. We're warehousing the data to solve problems, really. The data is just there so that we can say yes or no to a return so that we can help identify employees that are fraudulent. We actually have products that do marketing as well, where we're trying to figure out what incentives to give to individuals. And all that's really more of an analytics play. Got it.

So I wanted to talk a little bit about your book. And Reid and I both share books all of the time. And the book is titled Essentials of Modeling and Analytics, Retail Risk Management and Asset Protection. And from my perspective, not only is it a very good book if you're into data and analytics and asset protection but I think it's the only one that covers the vast sides of data analytics and modeling in relationship to risk in retail I kind of when I when I went through the book I read it about six weeks ago and went through unrelated to this call, was actually looking at it.

And one of the things that came up is integrating analytics into loss prevention teams. There's a section about that. And you talk about people. Can you share a little bit about what your thought process was with our listeners there? Yeah, I really wanted to. And I see an evolution of LP teams trying to go down this path on their own to some extent. And to get, you know, our team has 10 PhDs and nine more with master's degrees. Getting to that level takes a long, long time.

But to get started, I was trying to give the sense of the skill sets needed to just get started. And some people, their first instinct is, oh, I'll find a new grad student. We'll bring him in and he'll solve all our problems. But in reality, the very first analytics person you hire should not be the random grad student. You've got to hire somebody that has some business acumen and can do both. They need to be able to solve business problems and talk to humans and not just be math guys and also be able to do the math, too. So, you know, our take on it is you kind of need somebody with a lot of these skills. And so I listed out in the book a variety of skills.

There's five different skills there. And in your first hire, you want to have some level of all of those skills. They need to be able to communicate might not manage their own projects team lead etc and otherwise you're going to get what happens is people don't know how to manage analytics people and they have no idea what to call on them excuse my language but to say are they really progressing the way they need to progress are they solving the problem in the way they need to solve it? So you need kind of a proven entity as your first hire, somebody that's actually put analytics into work, not just a random student that's really never done it in the business world.

Especially for an LP team who, you know, I view LP people as very no-nonsense, smart thinkers, and they're practical and they want things that work. Whereas sometimes you get somebody right out of grad school and they're more living in the clouds. They haven't quite learned how to apply their craft yet to solve real problems. That's where I was heading with that. So, you know, just a quick personal story here is I guess it was about 10 years ago. I was given some data responsibility and pushed very aggressively to hire data scientists. So I wish your book was available for that. And I would say if the listeners on this call are AP professionals that are adding data analytics folks to their team, that one section, if I had it 10 years ago, would have made my life a lot easier.

So working through, I highly recommend it. I know what it was like. And I can tell you that I went through a round of hiring data scientists and all different people trying to figure out what we should do in-house versus through a third-party vendor. So kudos to you on the book, you and your colleague there. And I really think it was great. I'm going to turn it over to Reid for and your colleague there, and I really think it was great. I'm going to turn it over to Reid for some questions. Thanks, Tom. And, you know, I guess to follow up on that, that's a bonus to the book is thinking about how to use data and then maybe a little bit about,

well, this might be beyond what I know how to do, So I need to have somebody on my team that can do this. And now what am I looking for? What really, what skill sets? And like you say, you know, very often we're all in that position where we've got to bring somebody onto the team and we just don't know enough to know what kind of person and their capabilities that they should have, what their capabilities are.

So that's another bonus. I want to talk a little bit about, since we're talking about data, that's the world that you guys live in, we all do, but that you're mostly specialized in. And so how do you look at these things? I noticed that your team works on entitlement programs, for example. And I know that at UF we had a grant to work on SNAP and WIC fraud, which is essentially food stamps and so on,

and that there are these dynamics out there just like with any other program where there was fraud. And so Dave and Dan, what can you tell us a little bit about what your organization does to help people better understand the dynamics of what's going on to enable them to better address those problems? So we've done that a little with our safety division. And for those programs, the problem we're trying to solve there is fairly simple in the sense that we're trying to identify people that would be disqualified

because of their criminal activity or because they're actually in jail, not walking the streets. And so there we're trying to figure out what are the triggers that may disqualify them from those programs. Dan, you want to elaborate a little more? Yeah, I would just say, you know, we'll partner with agencies to basically identify these folks and help save money. So we'll basically, we'll look at the data, jail data, and try to identify these folks. Basically, it's a withdrawal. All right, well, that's good stuff. So I think at the end of the day, what we do at LPRC and on the UF team is try to understand human behavior. Good, bad, benign. Better predict what they're going to do, where they're going to do it, how they're going to do it, when they're going to do it, who they're going to do it with, and a little bit about their motivation and motives for doing it.

And it sounds like, and what I know about you all, your team and your different offerings, is that that's what you do. You're trying to help understand, help your customers understand externally what people are up to, where they're coming, and then help them use leveraging what you're feeding them and your skill set to reduce the problem, suppress the problem, displace the problem, and then the same thing on the internal side. But can you guys kind of talk about the behaviors you pick up on, the data that you get, how you use that, and how you make sense of that, not only for yourselves, but as you said, to get it back to the practitioners that need something that's very understandable and actionable.

Our goal generally, Reid, is to bundle up the data. And what we collect from retailers, that data is typically point of sale data and a bunch of add-on tables like store masters and item masters, employee masters, sometimes a customer database. And we're doing a number of things there, but we're really trying to create variables. But what I mean by variables is just summarizations of the data that tell something about the data. So usually we're operating on a unit. So the unit may be a customer or it may be an employee. And then a summary of

that data might be, you know, something simple about employee might be how many transactions have they processed today? How many had voids in them? How many times did they run a refund to themselves? So we're trying to summarize up that data into variables, which can be very understandable. We do the same thing on the consumer side. We look at, you know, how many refunds they do, how many purchases they make, how profitable they are to the organization. And there's ways to summarize up that data to give you those snapshots of that individual. detection models we're calculating probably three or four thousand variables and there'll be things you know there may be categories of variables that we calculate at different windowings of time they're they're just minutiae of you know we might look at things 15 minutes within 20 minutes of an activity so you look to see okay they ran a post void did they re-ring within five minutes

within 10 minutes 15 minutes we'll come up with varieties on that um and then it's really about predicting outcomes so the outcome may be they got terminated later that month or the the customer side it may be they got arrested for shoplifting or they got you know identified by the store lp team and they were taken in the back room and interviewed and we're really trying to predict those events from all those variables. That's really the main thing we're doing and how we're bundling up that data. All right, fantastic. So that's part of it. It sounds like right now the lion's share of the data that you collect, that you bundle,

and that you recode and use is coming from the retailer, particularly transactional data from the point of sale. What other information data are you all collecting, Dave, to make sense of the world that I might have missed that helps further inform you and your practitioners? It's mainly internal data from the retailer. We've gone beyond point of sale to some extent, and we're heading down the inventory path, trying to get data on inventory flows and out of the store. We're trying to collect as much of the case database as we can where we have times where they've been investigated. And with our exception reporting system, we're beginning to collect click-level data of the investigators so we know whether or not they're opening an exception, right? So if an exception comes back and says, hey, you should look at this, and they just immediately close it out and say I'm not interested without even opening it, that tells me something different than if they open it, which tells me something different than if they open it and then they place it into a case. Even if that case doesn't succeed, it tells me what they're interested in and what's catching their attention. And all those facts can be used as outcomes to model against.

All right, excellent. Can you guys provide sort of a real-world example? how, in this case, a retailer, use your data to impact a problem, whether it's an individual situation, would be helpful, and then also, you know, overall, how they're impacting across the enterprise. Yeah, our objective is to bundle the results up and hand them or spoon-feed them little chunks of information that are very pointed. So on the return side, for example, the retailer doesn't necessarily interact with the data, the variables, the model at all. All they know is we declined somebody or we approved somebody. Now, where they may learn about it is during the algorithm development process. So as we're developing the algorithm, we're going to talk them through the types of things we look at to make that decision. They're going to interact with us and say,

hey, I'm a big retailer and I have a very rich loyalty program and my top tier loyalty customers need to be treated with kid gloves. And so we may take that into account in our models and say, we actually will never decline those top tier customers. And so we interact with the retailer on broad stroke levels. They're not in there engineering the actual models. They're telling us what they want to accomplish with those models. And in the end, the model makes the decision and the algorithm makes the decision automatically. On the employee fraud side and exception reporting side, we provide a tool with secure where they can

really interact with the data. They can write their own queries. And there, it's more of a hands-on experience. So that one is probably more traditional exception reporting. With our employee fraud models, what we're trying to do is get that a little more in line with how we do return fraud, where we're detecting the employee that may be suspicious for some reason or another. And then we're playing that back to the retailer to say, hey, I think Joe Smith over in store 123 is doing something weird, something that may be fraudulent. You should take a look. And then we'll literally, as part of the modeling process, we'll play back specific things that drove that person to be suspicious in our minds. We'll say, hey, they ran this kind of transaction at this time and we'll bring up the actual transactions and we'll show them how they relate to other employees in the store, other employees in the district, et cetera.

employees in the district, et cetera, and say, look, they're five times the normal, or they ran 17 transactions where 90% of their items were line voided. The typical employee does that once a month. Those are the kind of things we're going to be playing back to the retailer, which allows them to kind of interact with the data. Excellent. So let me ask you another question, if I could. Could you go explain in a user-friendly way, we know what big data is, you know, and the term that people use are throwing around out there, and then what machine learning does, how that

helps you and others make sense of and better use data in a user-friendly way, Dave? Sure. Big data is really, you know, it's a buzzword like a lot of the buzzwords that came about, but there is a meaning behind it. What people were getting at was big data was more about pulling multiple data sources together. And so, for example, a big data play we might have is to pull our crime data, and we've talked about this and we've actually done a little bit of it. Pull our criminal justice database with all the arrests into the retail world to say, hey, is there something in the retail world where I could use that data to predict the other data? We've done it in the opioid space, too. We've brought the criminal data into the opioid prescription monitoring data.

prescription monitoring data. And we were able to predict actually the likelihood you're going to overdose on opioids is 7x greater if you've been in jail, for example, 14 times greater if you've been in jail for a drug-related offense. And so we've used that, that sort of big data in my mind is coming, bringing data together from multiple sources. Like if I bring point sale and inventory and crime data and all that, and I put it all together and I solve some problem with it. Now, the machine learning piece of it, machine learning is sort of the evolution of predictive modeling, really. Predictive modeling came out of statisticians and social scientists, and machine learning kind of came out of the computer scientist side. But they're really trying to address the same thing. They're trying to do the same thing. They take data, they turn it into variables, and use those

variables to predict an outcome. Now, the differentiator and what's got machine learning all the buzz lately is that it's been applied to some pretty cool problems like computer vision, speech detection, self-driving cars, and this has kind of gotten it a little bit more notoriety because you can enter you can see that the general public can see when alexa talks to me and answers my questions and looks up things for me uh that's cool that's like a new thing but that's really machine learning being applied to uh speech recognition now in the retail side, though, machine learning kind of comes to play is now we have complex problems we want to solve, and we're really just taking the things that we were solving with predictive modeling in the past, and we're sort of upgrading it a little bit.

And the cool thing about machine learning is it looks at very complex interactions automatically. So you don't need, I mean, you probably can solve the same exact problem with predictive modeling, but what machine learning gains you is it's going to automatically look at variable interactions that you can't, that the human may not have caught up, caught on to. And so that's really the next step. It's not much different than predictive modeling. And in fact, people that do predictive modeling now use machine learning techniques. And the fields have actually pretty much merged on top of each other. Now, the whole AI things of umbrella over all of it.

And that's usually it's got machine learning and predictive modeling, which are really mathematical techniques inherently. And that's kind of how I see everything relating together. I want to throw out another thing to you all, and that is we have what we call innovation chains here at LPRC. And so we, you know, we innovate to specific problems. This is a problem-solving community. So if we've got individuals or crews that are using sledgehammers to breach walls and go in and take Apple technology, steal them from stores or distribution centers. If we've got that's a specific issue, we're now going to try and define who, what, when, where, why, and how is driving that. And then we're going to say, okay, let's outline some specific solution sets that would address a would-be offender

or crew of offenders at each step or stage of committing, in their mind, successfully committing that crime. We're going to lay that out. We're then going to, okay, here's some tactics and techniques and technologies that we could use for step one as they ideate. Step two is a, you know, recon and initiate and so forth. Then at the next level, the next link in the chain, we're going to move from the lab environment where we can kind of do whatever we want to do in a way

to one or more locations, maybe in Gainesville, Florida, right here, where we can now adjust and dial in, adapt what that solution set is to that drugstore environment or a convenience store or department or specialty or whatever. Work together. How do we do this? Okay, now let's go to the third link. Now we're going to go, say, to this Baltimore region, three or four county area. That'll be our research lab as an ecosystem. But we'd like to look in zone five. We'd like to understand not only the data that are coming out of our point of sale and out of our item movement and other data sets, how can we look at who's taking advantage of those things? Let's say in Baltimore, do you all have any ideas, suggestions that you could share with this podcast audience on how you would look at that specific market in that area and understand the dynamics there and ways to better address them. So to reiterate your question, you're interested in focusing on a specific

crime type within a specific geography and how we go about building that up or kind of getting to what you call zone five? Is that your question, Reed? No, exactly, David. I think, you know, let's say in the Baltimore case, and we've got theft, a lot of it seems to be caused by and even driven by the ability to conduct illegal illicit returns in that dynamic? And let's say it's that Baltimore, greater Baltimore area, the three counties up there that we're interested in. Yes, is there something that big data could do to help us paint a picture as an individual retailer, but also in this case as a community of 16,

20 retail chains working together and working with law enforcement, all the agencies at the local, county, state, and federal levels. Yeah. From the machine learning predictive modeling side, you might look at if you can get your hands on actual events. Obviously, that's the best case scenario. And it's back to just the things you might put together to do a normal regression

analysis or to do some kind of analysis of an association between different things. So if you can say, okay, let's say I have a hundred stores participating in that pilot from, I don't know, 10 different retailers. And then you might say, okay, well, can I boil down what are the characteristics of those stores? And you could look at the likelihood they're going to be targeted or there was an event there. And then you're really just building models to relate those facts together. things you might be able to change in those stores to prevent or to limit the criminal activity there. Is that kind of what you're aiming for, Reid?

No, I think that's exactly it. We're looking, as you probably know, at robbery, parking lot violence, other violent events, but we're also now looking at maybe there are truck hijackings, there are issues in the supply chains that last mile delivery on the doorstep, and return fraud, what's going on with our refunds. We're trying to understand and map and measure and then affect and see what happens when we do that in a larger scale area and could big data and some of the machine learning and analysis techniques that are out there help us? We're working a lot with mapping now, but we want to continue to look for innovative ways to make sense of what's going on and then measure and observe what happens when we take action. If we implement some solutions in some test stores and not others, what happens?

What are the ripple effects? I mean, take your truck example. That's a great one. I mean, if a truck's robbed on the way to delivering to a specific store, I mean, you might, for example, that's a very specific problem. A lot of the machine learning and the predictive modeling tools are engineered to solve very specific problems where you can pinpoint an actual outcome. So there you may say, okay, well, how many robberies have we had in shipment to these stores? And then you start looking at variables about the robbery or about the shipment and say, okay, well, what about ones where they didn't have a robbery? Was there any differences there? For example, were the robberies more likely to occur if the shipment was after 6 p.m. or before 5 a.m.? Or, you know, kind of breaking down what are variables that are involved, like, you know, what was actually in the shipment?

Or what kind of truck was it? Or were there two guys on the crew versus one? was it or were there two guys on the crew versus one or you know kind of just breaking down each and every variable there and seeing how that relates to the probability of an incident or the probability of a shipment being stolen and i mean that's kind of how we always look at these problems is trying to figure out an outcome that you can measure and variables that you think just logically might predict that and then you just go about using a database method uh breaking that down and then the next level is once you can identify those sets of things you can feed that into machine learning type algorithms

to give you a little bit more edge in terms of it'll pick up on more complex interactions like you know if it's after a certain time at night and there's you know this is the particular shipment then i have a way more likelihood of a theft occurring than, you know, if you just looked at one variable at a time. Excellent. Sounds good. Tom, I'm going to go ahead and toss back to you and for more insight and questions. Yeah. So, you know, David, you talked a little bit about BI and big data. Can you give the listeners kind of a layman's term in the simplest terms, what artificial intelligence is in relationship to that? Yeah, I think the simplest would be it's something that mimics a human action. Right. So a dumb one might be like I always like an ATM is kind of a dumb example. A dumb one might be like I always like an ATM is kind of a dumb example. Right. It's not really actually new. Newer ATMs are.

But originally you typed in your you put in your bank account information and you slid your card in and money came out that that's automating a human task. Right. It's very it. But it's not that intelligent. But now it scans checks and it and it's probably looking at your face when you're standing there too to make sure you're you. There's probably a lot of other things you don't even see that's going on. That's sort of basic AI. It's taking a human task and making a machine do it. Now, everyone thinks of AI like it's a robot and it's going to talk to me and learn anything. That's what's called

general AI. General AI is what everyone's scared of. That's where you invent something that can literally do exactly what humans do, learn like humans do, act like humans do, and yet has unlimited databases of knowledge and all this stuff. That's what people, that's what you get Elon Musk saying is, that's the singularity, that's all that stuff. That's what people, you know, that's what you get Elon Musk saying is, you know, that's the singularity. That's all that stuff. What everyone's talking about today, what I'm talking about at least is what's called narrow AI. And that narrow AI is more about solving a specific problem. And, you know, whether it's playing chess or, you know, the game go,

that's a very specific problem or even Alexa is a very specific thing. It's right. It's, it's, it's engineered to sit problem. Or even Alexa is a very specific thing. It's right. It's engineered to sit on your table and listen to your voice for certain commands and answer questions. Siri does the same thing. Now, what we do is kind of this invisible AI. You can't even really, you don't even really know what's going on, right? It's taking in data and then it's just saying yes or no on a return.

Or we're saying, saying hey look at this employee they might be suspicious it's not listening to anybody's voice it's not using computer vision it's but it is using an enormous amount of data and doing things automatically that maybe a human might have done historically um but it's trying to do that using a broader set of data. That's sort of my simple answer on AI. It's not, and there's all ranges of AI. The self-driving car is probably like the coolest, most visible AI thing that's come out lately, but there's a lot more going on under the hood and have been for 30, 40 years.

And I know when you've been speaking a lot about machine learning, I'd ask, can you kind of explain the relationship between AI and machine learning to the audience as well? Because I think you mentioned before big data is a buzzword. AI becomes an all-encompassing word, and your narrative and example is probably the best I've heard where it's sobering and says, hey, it's mimicking human behavior. Can you just give us kind of a brief overview of the relationship between machine learning and AI and what you're actually doing? Sure. Machine learning is, and if you go get a book on machine learning, one of the first things in there is going to be regression analysis or logistic regression. one of the first things in there is going to be regression analysis or logistic regression. And so it's really machine learning is an evolution of what historically has been called predictive modeling or just modeling.

It's taking a bunch of variables and trying to predict an outcome. That's the essentials of predictive modeling, of general modeling, of regression analysis, logistic regression. Now you have the more fancy math in there. You have things like extreme gradient boosting, deep learning neural networks, but they're all trying to do the same thing that the predictive modeling had tried to do all along. They try to take inputs or variables and predict an outcome, whether it be, hey, I'm standing here and I, I see 30 things going on. Can I predict what's the likelihood someone's going to steal something right now?

You know, I got a guy, he's got a hoodie on with dark sunglasses and he's been milling around my most expensive aisle. Okay. Now I've taken like five variables or 10 variables there. And I'm saying just logically in my head, what's the probability now a theft occurs on that aisle? That's inherently your brain doing a predictive modeling exercise. But you can set all that up. As long as you have enough data, a model can do that. That's really the connection. Now, AI and machine learning, machine learning is a subset of AI.

It's a tool. Think of AI like a toolbox or even the self-driving car. Self-driving cars is a great example. It has lots of pieces in it that are machine learning based. So the computer vision piece, you know, there's other pieces where it's looking at sensors in the bumpers. There's, you know, probably five layers of machine learning in there that are trying to solve different problems and telling that car whether to go right, whether to go left, and they're trying to avoid, you know, accidents. They're trying to stay on the road, stay between the lines. There's all sorts of outcomes they can predict and measure there within that body.

Now, within like a returns management platform like we have for Verify, our machine learning is really simple. All it does is take a bunch of data on the consumer, maybe on the employee, maybe on the store, and it's saying, what's the probability of fraud or not fraud? Or what's the probability this consumer makes 30 more returns in the next year, but no purchases? That's the sort of stuff we have an outcome to predict. And all those relate together in an AI system. So an AI might be just something that solves a specific problem, whether it's driving a car or doing an ATM machine for getting bank transactions or saying yes or no on a return transaction. The AI is the whole wrapper. It's all the rules. It's the software. It's

everything. Whereas the machine learning would be one algorithm within there that's solving one very specific problem. Hope that helps. And I would add, just to kind of piggyback on that, statistical modeling, traditional statistics, and machine learning, the objective of both of them is just to learn from the data. And as it is implied with machine learning, there's a lot less human interaction involved in contrast to traditional statistics where a statistician might use data reduction techniques, his assumptions about the data and its distributions,

whereas with machine learning, machines are exposed to the data, they learn from the data, and like David said, they get smarter in the student-packed loop. So that's kind of a little bit of the difference between the two. There's a lot more human interaction with traditional statistics. Great. Thank you. No, that definitely answers the question. I think it's probably the number one thing that comes up for me when people ask me off the side of, you know, what's machine learning, what's artificial intelligence, and, David, Daniel, the NRF published a short overview of the top tech investments right now in retail. And not surprisingly, advanced data analytics and artificial intelligence were both listed in the top five.

Just curious if anybody had a chance to read that and any points of view on the advanced data analytics piece of it and really explaining, you know, what the difference between data analytics and advanced data analytics would be to the user with some, you know, a real world example. Yeah, I view just data analytics more like reporting, whereas maybe advanced data analytics, you start to get into predictive modeling machine learning type stuff so i the difference between advanced analytics and ai i i mean that's just i i honestly don't know it ai is an advanced analytics in and of itself um you know you a lot of companies are going to go out there and propose report ai i have ai or i have this. The first thing I look at to see if they're serious is if you don't have PhDs on staff or guys with master's degrees or advanced degrees in a data science type background field, then they're just spinning. They're not really telling you the truth. So I think, you know, I looked at that article and I saw it and people are definitely trying to jump on the bandwagon and capitalize on this being a top thing. But, you know, there's people seriously doing it and there's people that say they're doing it.

And the big differentiator there is really what tech talent you have on staff. As far as the guy's predictions about AI and what people invest in, he kind of implying, hey, you better get ready to get into AI. Make sure the juice is worth the squeeze. Well, I think LP professionals have been doing that for years, right? If you deploy any solution, I don't care what it is, whether it's a sensor or an AI solution, you're going to test it and make sure it works. And so as long as you're going through with that process, there's nothing to be scared of here. You just, because usually even, even though the under the hood part of AI is a little bit scary,

you may not know what's going on. The outcome's usually pretty simple. And if the outcome's not really simple, then you probably need to stay away from it because it's, it's usually, I'm going to find a bad employer. I'm gonna find a bad consumer. I'm going to figure out who to give an offer to if it's a marketing campaign. And those things are very simple problems to articulate. Now, how you solve them and how you bring data together, maybe a million lines CFO and say, hey, if I deploy this technology, I tested it and I'm going to I'm going to gain, you know, I'm going to decrease shrink by 10%. Like that's the measure you should be employing. Don't worry about what the technology is.

But it's data analytics is definitely a hot topic and there's some cool problems to be solved with it. We're kind of a testament to that. But the way that LP professionals should look at it is, how can I measure it? And did it really work for me? And that's all you need to do. I think as data continues to grow, you know, we'll probably see a cultural shift. But AI is not going to be a panacea. And I think, you know, people are going to start integrating it as maybe part of a larger ecosystem.

But like David said earlier, I think you need to get a really good team lead, start slow, kind of get a roadmap and develop a strategy on how to use it. Make sure you have someone at the top who has a big picture idea and also can get in the weeds and understand what's happening analytically.

Right now, a lot of people are using it in kind of customer retention, recommendations for, you know, products for shopping. But I think we'll slowly start seeing a cultural shift with, you know, LP starting to use more advanced analytics. So this has been fantastic. I really appreciate you all joining us today. So, again, APRIS doctors David Spites and Daniel Downs, Control Tech, my partner in crime, Tom Meehan.

And I want to thank, as always, our producer of the Crime Science Podcast, Kevin Tran. Thank you again, everybody. And we look forward to meeting up again. Thank you. Thanks, Reed. Thank you, everyone, for tuning in to this episode of Crime Science. We also want to thank Bosh again for making this podcast possible.

If you would like to suggest topics for future episodes or provide feedback, please email kevin at lpresearch.org. See you next time.