We humans are, as a whole, a predictable bunch. We like eating around the same times every day, taking the same routes to work, going to the same restaurants and grocery stores, seeing the same people. Yes, humankind has an inbuilt sense of adventure and a desire to explore the unknown. But day-to-day, we’re largely creatures of habit — and machine learning and artificial intelligence are trained to learn those habits.
The Covid-19 pandemic took us from creatures of habit to creatures of necessity. Our patterns shifted nearly overnight, and our routine-driven lives quickly became a game of calibrations: What do I feel safe doing? Who do I feel safe seeing? Where do I feel safe going? No industry that depended on A.I. or machine learning to drive its customer experiences was immune to this influx of aberrative data from customers; like all of us, the algorithms were simply thrown for a loop.
“Covid was a shock for people, but it was also a shock for machine learning algorithms,” says Joan Bruna, an assistant professor at the Courant Institute and the Center for Data Science at New York University who concentrates on machine learning and deep learning topics. “Applications that relied on human activity underwent a severe overnight shift. A bunch of our activities changed and the algorithms had to change in tandem.”
But flexibility in the face of crisis is something that humans deal incredibly well with, so when Covid turned the world — and the algorithms it depends on — upside down, engineers got to work transforming models to reflect our new reality.
“When you build models, you make many necessary assumptions about the data and the process you’re trying to capture, then something like the pandemic comes along and breaks those assumptions,” says Bayan Bruss, Director of Applied Machine Learning at Capital One. Bruss and his team work with data scientists and engineers across Capital One who are applying machine learning to solve important problems such as enhancing digital banking experiences and fighting fraud, as well as researching shifting customer habits in the wake of the Covid-19 pandemic. “The universe is an ever-changing place so models should have the ability to adapt.”
The Shift from Stationarity
“The basic underlying principle of machine learning projects is stationarity — a core assumption that data doesn’t change,” says Bruss. “When you train the model, you make the assumption that the relationship between features, model and outcomes isn’t going to change over time. If you talk to data scientists, they know this assumption is false, because the world is a dynamic place. However, most changes in data are gradual, so models have time to adjust.”
Needless to say, the rapid onset of the pandemic didn’t leave us (or models) much time to adjust. One crucial development in consumer behavior that happened nearly overnight was the overwhelming shift towards e-commerce. Online shopping has been steadily chipping away at retail market share over the last decade, growing from 5.6 percent of total retail spending in 2007 to 16 percent in 2019. Covid-19 has only accelerated that trend; with all but essential businesses shuttered for months and social distancing measures in place indefinitely, many have moved even their most basic transactions online.
These new consumption habits have created corresponding changes in the patterns machine learning algorithms learn from. “If you build a model that is based on someone doing all their grocery shopping in person and now they’re doing all of it online, it looks fraudulent,” says Bruss. “But when the whole population changes its behavior, the entire system starts to say, ‘Hey, something else is going on here.’”
The New “Neighborhoods”
That’s exactly what happened when businesses started to close due to Covid-19. Pre-pandemic, much of our consumer habits were based on goods and services being close together physically. You might go to a hardware store and then a coffee shop because they’re located next door to each other. But during Covid, our habits have been clustered around service similarity because that’s typically how we shop online, according to research from Bruss. The new transactional “neighborhoods” have given engineers and researchers insight into how exactly Covid has transformed the way we shop.
If it seems a bit heady, consider the utility of being able to understand and rapidly adapt to new customer habits and better serve their needs. “The data you have and how you harness it is critical,” says Bruss. “At Capital One, we use our data to make the customer experience better. We learn how to best solve customers problems across our digital experiences and in our call centers.” Data around our “new normal” shopping patterns can also be used to retrain fraud-detection models so they’re able to more accurately discern the difference between truly fraudulent activity, and the sort of larger-scale purchasing shifts we’re all undergoing.
Adapting to the New Normal
It’s safe to say 2020 has brought radical, and rapid changes in the data set of our behavior, and tested machine learning algorithms like never before. As a result, it’s given us our first true test of the adaptability of our machine learning systems at a global scale.
According to Bruss, the key to making machine-learning models as flexible and resilient as possible is to “implement continuous monitoring of the variables that go into your model and the predictions it makes. Monitoring means regularly analyzing both inputs and outputs with a variety of methods to determine if something has changed.”
Bruss’ research also aims to train train modes to go beyond pattern recognition, and become capable of making “causal inference” — that is, capable of understanding the causal relationships in the phenomena you’re trying to model. “Knowing causal relationships allows you to know what to do when something does change — change your policy, refit your model, et cetera, and the effect you’d expect that change to have,” says Bruss.
Understanding that two factors are related, and making real-time adjustments accordingly, is the sort of thinking we humans do, well, without thinking — for example, we might see that it’s snowing, and leave earlier for work. Because we understand there’s a causal relationship between weather and road conditions, we’re able to infer that we’ll need more commute time.
Looking Ahead
The fields of machine learning and artificial intelligence are poised to grow as researchers continue to unlock their potential. Bruna sees one potential area of expansion in “the application of Machine Learning algorithms to computational science, such as chemistry, biology or physics.
For now, Bruna says, “machine learning has become a vast ecosystem with a wide range of applications. My recommendation to aspiring machine learning scientists and engineers is to aim to balance breadth and depth in their machine learning skills. In other words,” he says, “do not worry about mastering every aspect of machine learning, but rather be selective in a domain and spend some time and effort learning the fundamentals behind the algorithms.”
It seems certain that businesses will increasingly embrace artificial intelligence’s potential, too. Fundamentally, better trained algorithms can help a business better understand, and better serve its customers.
“Machine learning allows us to unlock a host of new opportunities to customize the user experience,” Bruss says. “What we’ve been able to do from machine learning is identify customer intent, give them the information they need more rapidly, and help people in a way that’s accessible, fully customizable and dynamic.” In a time of widespread uncertainty, that mission feels more important than ever.