Using machine learning to change banking for good

From informing customers about unusual charges to answering their questions in real time, Capital One is using machine learning to bring simplicity and humanity to banking. And our efforts with machine learning are just getting started.

We recently spoke to Abhijit Bose, Managing Vice President of Enterprise, Data and Machine Learning and Head of The Center for Machine Learning, about what makes Capital One a top employer and how the company is using machine learning to improve customers’ experiences.

How did you get into machine learning?

I didn’t get involved in machine learning until I wrapped up studies in another field. I was working at Nokia Bell Labs, a scientific development company, after completing my Ph.D. in engineering mechanics. Nokia was where I first saw machine learning. I was fascinated by a field in which you could predict the future for all kinds of things. It was this whole other world that I didn’t know existed.

I returned to school to earn a second Ph.D. in computer science and engineering. Over the next couple of decades, I worked in machine learning and artificial intelligence for companies like IBM, American Express, JP Morgan Chase, Google and Facebook. I love the work because machine learning goes through rapid changes. I wanted to be part of something that’s always adapting.

What made you want to come to Capital One?

Three things about Capital One appealed to me.

The first is that we do banking but are not just a traditional bank. We are a technology company at our core, focused on innovating the financial industry. Capital One seemed like a great place to work on challenging new data problems and advanced data-driven science. We have the autonomy to create a machine learning platform that is cutting edge.

Second, we are on Amazon Web Services, a cloud computing platform. We don’t work on legacy platforms. I know all the headaches that come with working on outdated systems and we don’t have to worry about that at Capital One. Here, machine learning is done with first-class infrastructure and in a data-driven environment on the cloud.

Third, many engineers in The Center for Machine Learning publish research and both attend and present at conferences. We have contributed patents in data science and computer science. We’re continuously growing and adding to the field through our work here.

What would you want a candidate to know about machine learning at Capital One?

Capital One is already quite advanced because of the history of our analytics. Since its founding, data has been at the heart of Capital One. We believe in the power of data to drive insights and empower people to deliver real-time solutions to millions of customers, making banking simple and accessible.

Frankly, machine learning and data are now two important pillars of the company. There is room for engineers and scientists to join Capital One and explore new machine learning techniques and neural nets. We also need product managers to come in and help with this transformation.

The work we’re doing is purposeful and meaningful and has a huge effect on customers and the community. Financial life is so important for everyone. Decisions we’ll be making in machine learning have a lot of impact on our customers, whether it’s a model or a platform. Your work makes a difference here.

And once you arrive, we have the tools and workplace culture to help you succeed and perform cutting-edge work.

Tech College is a learning platform created in-house that helps all associates learn new skills and further develop their tech knowledge. The platform offers classes that provide technical skill development across machine learning and artificial intelligence. We also partner with some great companies, like Coursera, and leading industry experts to design the Tech College curriculum.

Finally, no matter who you are, you will be welcomed and respected. You will be working among a team of talented, diverse people.

What makes someone successful in machine learning?

You must constantly learn. I am always learning. Whatever was true in machine learning one to two years ago is not working anymore. This field goes through rapid churn. The platforms are getting newer and newer, including open-source capabilities. You can get so focused on building something that we forget to look around and see what other innovations companies and people are producing. In machine learning, you have to be curious and observant.

The other thing is people have to explain complicated topics in a simple way. A machine learning expert must explain what they do and what they’re working on to a broader audience—business analysts, product managers, designers, analytics people or sometimes engineers. It just creates confusion and a bottleneck when people can’t articulate updates or accomplishments.

How does machine learning add value to our customers?

One way we use it is to fight fraud. Machine learning uses data to help identify fraudulent transactions on your account and alert you to the problem in real time so you can take action. We have your back.

Another way we use machine learning is by offering our products to customers. Machine learning helps us understand our customers. If you had a fraud attempt and you are using the Capital One app, that’s not the right time for us to offer a credit card to you. We use data and decisioning to ensure you’re getting messages that apply to where you are in life.

Where do you see Capital One’s Center For Machine Learning function going next?

I’d say we’re about 10 percent done, so we have a big future ahead of us and questions to answer. We are building our foundation to deploy more advanced neural nets and gradient-boosted machines models.

A lot of work is going on behind the scenes to ensure what we're doing is understandable for the general public. The traditional models we use today have very good explainability. We need to ensure that everybody understands what we're doing as we advance in machine learning. That builds trust.

Of course, we want to create responsible AI. How do you detect bias in models? How can we detect bias in data? For example, you could abstract 1,000 customers from our database and get a disproportionate representation from one gender. As a result, you’re building a model that has a huge impact down the line. The data doesn’t represent the entire population. We have to figure out how to have responsible AI and data so our results do not skew.

No matter our next project's focus, we’ll continue to be a company where associates are empowered to be their best, challenge themselves and disrupt the status quo. They can be leaders in machine learning.