Welcome to NVIDIA's AI Podcast. We're recording from the floor of the 2017 GPU Technology Conference a gathering of the AI faithful here in Silicon Valley.
If there was an industry that AI seems like a perfect fit, it's payments.
Training a deep learning system to sniff out fraudsters is the obvious and perhaps most critical application for a payments company and its application of AI.
But there's a lot more. PayPal has been in the AI business for a long time, and it's certainly used AI to help detect and combat fraudsters. but there's a lot more.
And to help us talk about that, we have Vadim Kutsi, Distinguished Architect and Data Science Lead at PayPal.
Vadim, welcome. Thank you. Thank you for the introduction.
It's great to be here. I've been spending my second day here enjoying the conference and enjoying great things which have been presented in the conference and talked about.
And as you mentioned, PayPal being a part of AI from the beginning, it's in our core, even before PayPal become known as a PayPal, used to known under other names.
A big part of what was at the core of the company.
How did you guys come to it so early? And when you say it was already there for some time, how long ago are we talking?
We're talking about end of 90s when PayPal was just created and Truly to say, we did not come to it because we wanted to.
We came to there because life made us to do it.
If you think back of 90s, Wild West, on the web, and as soon as there was a payment company where you can take something, money to send to somebody else, they were not very good. players in the system who start playing alone and as PayPal were realizing how much money they were losing on those bad players. we start looking into the models and we start looking into the data, how it can help us.
Whenever you talk about fraud, there are always some very easy steps you can start with.
Yes, you can check credit card or you can hold processing for three days, but Customers do not like it.
Going back, when people start looking into fraud more and more, if you even think about beginning of 2000, The most common way to promote the front was to block particular countries.
Big Nigeria, pink Romania, it was most obvious place to do it.
But guess what? Those countries also have business and you also want to support those countries.
So, blankly removing the country or blocking set of IPs would not work because you're blocking good customers and you're removing their experience.
So you guys had to use machine learning, and I don't even know what you called it back then, but... to distinguish the good customers from the bad customers.
Correct. And how did you start going about doing that?
You guys had the advantage and you still have the advantage because you guys had the data.
So how do you go about... you know, setting up a system that will detect, you know, good customers from bad.
So again, if you're thinking about 15 years ago, a lot of banks were doing something very similar.
If you look at the very large banks, they would have their fraud department, which would create their logistic models and based on their logistic models would try to prevent the fraud.
The biggest difference between what we did and what banks did, we collected our own data.
And based on our own data, we would build deeper models.
And that was really into what kind of data we created.
And yes, when we were doing AI 20 years ago, it most likely was logistic regression, which worked great.
And even neural networks, deep learning, It's not, they were discovered back in 1980.
There was simply no power or no computing, no enough data to do anything with it.
And even today, if you can build a model where logistic regression works, I don't think it would make sense to do anything else.
Now, today with more and more use cases for AI, in most of the cases, logistic regression would not work and that's the reason you start looking into other models and other places, but not because you just want to replace legislative regression since it's old or since it's been used forever.
It's been used forever because it works.
So is AI used in different parts of PayPal for different kind of attacking different fraud in different ways or different actors.
I mean, if logistical regression works great for those bad players that you described at the beginning, Has the universe of its use expanded within PayPal?
Even legislation still be used in PayPal, we use a lot more models, and we use Deep Learning and any other models, a lot of different models there.
And yes, the use of those models have highly expanded within PayPal.
Even if you look within a fraud, If in the beginning we were just trying to stop transaction, now we're looking into merchants.
You know, there are bad merchants, there are bad player merchants who wants to play the system.
They're merchants who sell products which we do not want to support or it would be illegal for us to support, and we're using AI technology to sort out those merchants, even before they start making transactions.
The frauds have become a lot more sophisticated.
Sometimes they might want to use network of the particular financial institutions not to do fraud, but to prepare to do fraud somewhere else.
Let me give you a very basic use case, which was very popular about 10 years ago.
In the restaurant somewhere, the fraudsters would swipe credit cards of the customer and they would have numbers of credit cards.
What they don't have is they don't have zip codes.
But taking a guess, there are five zip codes in the area, so the credit card probably would be from one of those five zip codes.
What is the best way to try to check zip code for the credit card?
You do authorization, just basic authorization, and see whether authorization goes through or not.
So from the perspective of the company who does authorization, the company would not be losing money, and it's not a fraud from the perspective of the company.
However, preventing cases like that would stop future fraud and would help people and consumer who eventually becoming consumer of the company.
And PayPal been doing and working closely with authorities on trying to prevent the future fraud.
But even if you go outside the fraud, artificial intelligence and machine learning can be used in a lot of different places and one of them is operations.
Operations and by operations you're describing what?
By operations I mean computer, our data centers. and our apps which are in Consolidated Data Centers.
Right, so fraud, certainly AI is deep in there.
Security. Yeah, security fraud is our security is there.
But operations, the actual rank of our computers, we are starting to use more and more machine learning and artificial intelligence to make it better to make our system more reliable and to make sure that we serve our customers better.
So how is AI applied there? And again, what are the problems you're trying to solve?
And what are you training this AI to do?
So if you think about operations for the company size of PayPal, so now PayPal is a financial network.
And number one thing which PayPal wants to make sure about their systems is that they're available at all of the time.
You as a consumer, if you want to make a payment, the last thing you want to do is to get an error so your payment didn't go through because of the error. with the complexity of the systems.
And PayPal is not unique. You take any other payment provider, systems are very complex.
System goes through a number of different checks and balances.
If one of the applications or one of the small node in all the chain fails, it might block the transaction.
Right. So we need to monitor thousands and thousands and thousands of nodes applications, VMs, containers, all of the methodology you can think of on how they perform.
And as soon as one of them fails, try to figure out why do they fail.
Now, this is interesting because in your case, you describe a transaction and transactions are how you guys make money.
And so your operations need to enable those transactions just 24-7, right?
But it also implies, I mean, I think it shows a path for other companies that, hey, your business relies on your operations and your data center and increasingly your data center within a company, And you need to make sure that that's running because that's what your business relies on.
So I just think it's interesting that you can draw the line more between AI and, okay, here's how you apply it to business.
You're exactly right on that. So, and it's not only transactions, it's also customer experience.
We want to make sure that when customer wants to come to PayPal for any reason, just to send a quick note on the Venmo feed about what they're doing. it's up there, it's running.
If it would be down, if something could happen, it's bad customer experience.
So running the network, making sure It's available.
It's not only enabling the business, but it also gives us ability to run it more efficiently.
When you look into creating more reliable network, there are a few things you can always do.
Number one, you can put a lot more hardware on it and make sure that hardware runs at much lower rate.
You can also buy more data centers and make sure that data transactions or any kind of operations will go through a number of different pools located in different data centers. all that cost money.
So by writing AI, we are monitoring our system more efficiently, and we are enabling our business not to buy as much hardware and not to spend as much money on the operation.
How has it been to kind of convince those who would be on the side of, okay, let's just buy more redundancy?
I mean, how has the evolution gone where they've learned to trust what you guys have done.
You know, it's a very interesting question.
A long time ago, I used to work for a company called General Motors.
It was back in the 90s. And back in the 90s, General Motors was as far from Silicon Valley hype from all of the stuff we are talking here.
And in General Motors, It was very small problem.
There were people running operations as a plan.
There were robots and there were workers helping those robots to run it more smoothly.
And my job back then was trying to optimize maintenance for those robots. to make sure that the line on the plane does not stop.
Right, and we're talking building cars. And we are talking about physical building of the cars.
And it's also was very, very similar because when you talk to operators at the plant who are mostly union workers working there on the shift and paid by hour.
They're always they're always inclined of having two robots doing same stuff or having two separate conveyor that one stop second would work.
And when you go back to them and say no, by having better maintenance on this robot, and having that maintenance done every three weeks versus every four weeks, or having that maintenance done when this particular temperature sensor goes outside, or even more,
Having maintenance on this robot done when down the line we see slowdown because there is a dependency would save time.
And we can do maintenance at night. the people would start believing you after they see results.
It doesn't happen overnight. It doesn't happen over a few weeks.
You start with something small, you pick up something small, where you can show the result, you start working on that place, and when they see results, they let you expand.
Over my career, I work for General Motors, Automotive, I work for Biotech, I work for a very small startup, and I work in a number of different places, eBay.
And prior to it, I even worked for Bell Labs doing some mechanical manufacturing.
And operations people are very similar across all of those industries.
They're all the same. their job is to make sure that operation runs.
And they don't believe anybody else outside what they do because that's then what they do and i completely understand it and getting their trust it always takes time right well that is a a point that we should all take to heart I think because one of the issues is there's people like you who understand AI and are you know all in and all behind it and then but you still need to sell it outside of your team or your organization more broadly within a company or It depends on what organization you're a part of.
Pardon the interruption, but we have a favor to ask.
Leave us a review on iTunes, Google Play Music or whatever your favorite podcast platform of choice is.
It helps more people find us. As always, thanks for listening.
Now, back to the good stuff. You mentioned eBay.
I know eBay uses AI in different ways, and it's certainly not the same way that PayPal does.
How did eBay... deploy AI and how would we as eBay customers see its impact?
Let me talk about time when eBay and PayPal were together.
There was about four years ago. And at this time, eBay also owned another company, which back then used to call it eBay Enterprise.
And that company would run large merchant websites.
And eBay decided to use AI to help those merchants.
So you as a customer of one of those large merchants, when you come to the merchant website, you probably don't even know that it was run by eBay. or that software was written by eBay.
But your help to better navigate the website and to have better experience on that website.
And one of the big things which we did for those website was customization for the customers are known to the merchant.
It's very common that if you shop on the same merchant over and over and over again, and let me pick Toys R Us.
If you buy from Toys R Us diapers every month, I'm pretty sure 12 months down the road, Toys R Us can guess that you come over to buy diapers.
Maybe by the time you were not buying diapers, but Toys R Us was still gas.
Now, when you're the first time coming to Toys R Us, Toys R Us have no idea what to show. what they a lot of large merchants what they do they would pick four and five main categories and they will create more or less slideshow on the front page And you have to sit there and wait until your particular slideshow show up because it's probably have some kind of coupon or promotion.
Right. eBay started building artificial intelligence back then.
It was neural network-based model. It was modification of neural net. in order to create a model which would be deployed to the merchant website and would do prediction for the new customers.
The model would be self-learning. So as more and more customers shows up, and based on where they go, the model would update itself.
But even more, something happens, there is new advertisement happens about the merchants or the new product comes out, model would learn about itself and it would deviate from whatever was before and start predicting better behavior.
The merchant used that technology for both personalization and for advertisement.
If a merchant particularly wants to advertise one particular type of the business, and let me, for this example, pick a merchant called Spanx, Spanx is producing shapewear.
One of the line of shapewear they're producing is men's shapewear.
Oh, well, they do Mensu Spanx, for those who don't know, is a kind of a hosiery The underwear thing?
Shapewear. They produce shapewear. Okay.
So you put it on, you look great. You put it on, you look great.
Okay. And it's mostly for women, but they do also have men's line of shaver.
And when we started working with Spanx, I was not aware of it.
They did not want to put advertisement for men's shapewear in general site.
They want to try to put it for the correct people because they thought it was important for their business.
We ourselves did not believe that our model would be able to predict it because We don't have gender as part of the models.
We don't have any kind of demographic information.
We just have people's behavior. But model was able to figure out to hook which customers to show advertisement for Mineshafer. at the rate that, again, neither us, neither Spanx believed it.
I still amazed by what Madel did. I have no clue how model did it because we put a lot of variables into the model, but we were able to predict it in Spanx. as a merchant was able to serve their customer much better and specifically advertise men's shapewear to the customer who would want to buy it.
Wow. Well, I have to know, what were the customers, like what was the profile that wanted to buy it?
I'm a data scientist. In my heart, I do data.
And as somebody a long time ago asked, we need to predict Can you build a prediction for moms because we want to sell some product for moms?
And my answer was, no, I cannot predict moms.
I can predict who would buy the product.
But whether they're moms or pops or grandparents, I don't know.
And even more, I probably don't want to know.
So in case of Spanx, they did ask us question about demographic.
We did try to look into demographic because you can purchase demographic from third party providers.
Everybody do it. There was no clear demographics of what we could pinpoint.
So there was no answer back then. And yet the model did a good job of picking the right people.
That's amazing. So back to payments, we live in a world now where I think online payments and digital payments are more and more commonplace.
Bitcoin recently went back through the roof again.
You know, I don't know if it's more mainstream than it was, but how do you see it being applied not just by PayPal, but more broadly within the industry?
So number one always will be fraud. How do we minimize the fraud?
And by minimizing fraud, it's not necessary how do we decrease rate of the fraud, but how do we decrease rates of false negative when companies think it's a fraud, but it's a good transaction.
That goes back to the good user experience of the customer.
When you use your credit card in the store and credit card got declined, simply because two minutes ago you were in another store, which listed in another city, but happened to be across the street.
You are not a happy customer. I know I was there about a month ago in that situation.
And that explanation did not click with me that, oh, you're in a different city, you're in Las Gatas, and five minutes ago you hit a transaction in San Jose, which is literally three buildings away.
The same goes for any kind of online payment system.
We want to make customer experience better and better.
We want to start taking more and more different type of currency.
Bitcoin is a good example, but it's by definition, Bitcoin will probably have higher fraud rate.
So what kind of models can we put and how much artificial intelligence do we put there?
But it's also not only in fraud, it's also in customer experience. when you come to the provider and you try to pay very often we can based on your behavior We have a very good idea whether you want to pay with a credit card or you want to pay with a bank account. or whether you want to use one credit card in one type of merchants and different credit cards in different type of merchants because you do get different types of points.
From the financial institution perspective, We probably can predict it.
Not always, but there is a good chance. And can financial institution provide that information?
When you go to the store and pay with your Apple Pay or Google Pay, and PayPal is going to be available on Google Pay soon.
When you come and pay in the store, can we tell you what is your preferred method?
And even more, how we can tell you that here is a better combination.
If you're going to be using your Chase card in this particular store, you will get more points.
Right. That goes back and using points on the credit card is just a basic example.
There is a lot more to it, but it's going back to the customer experience.
How can payment providers create better customer experience, make the customer transaction much smoother? and make customer happier.
Yeah, and not so that you're just buying more stuff or spending more money, but spending more wisely, right?
Or efficiently? From any kind of merchant, from any kind of payment providers, spending more money is always good.
Right. But we all understand that total amount of money spent probably more depending on the current state of the economy than anything else.
And it's overall, if you look at the bigger picture, it's probably total money spent.
None of the payment provider can change.
What we can change, we can change customer experience with that particular payment provider and using that, increase share of that payment provider.
Look at Visa as an example. 20 years ago when Visa started, it was a very small part of the total wallet.
But with this experience, that you just come, swipe your card, done deal, walk out. was so great that customers started adapting more and more and more.
Not because of the great advertisement visa date, not because of something else, but because of the simplicity.
And today, as we're talking about digital payment providers, it goes back the same way.
How we can make user experience easier. And I do believe making the user experience easier and better is AI is a big part of it.
Because based on the previous behavior, AI can predict that this probably is not a fraudster and we should not block transaction.
But even more, What else can we do with the user?
I look forward to all of that because I think we all rely upon our payments systems, whatever they may be, whatever our favorite may be in more and more ways.
Security kind of underpins it all, but we do want the convenience to use it how we see fit.
So Vadim Kutsi, thanks for joining us on the podcast.
And we'll keep an eye on what PayPal is up to next.
It's my pleasure to be here. Thank you.
