Capital One shortens the machine-learning curve
When it comes to machine learning, it's the human learning curve that can be most crucial to the success of a project.
At the FinDEVr New York conference in March, Sandeep Sood, vice president of software engineering at Capital One, shared an interesting case study on how his company uses machine learning to analyze clickstream data. His pragmatic approach to implementing the new technology can be viewed as a template for others.
The most useful lessons he learned in the process had to do with strategizing ahead of time (think big for this part) and then rolling out the first use case (it helps to think small here).
To illustrate his point, Sood went all the way back to when electricity became available for general use. The business world had run without it and figuring out how to use it took decades.
"There are tons of stories of companies that actually set up lab groups — the way we set up lab groups to explore machine learning or voice OS or whatever it might be — to figure out how to use this new technology called electricity," Sood said.
Many of the lab groups failed — "they would try these prototypes out and they couldn't take advantage of electricity in the way they thought they could" — because their focus was on improving the existing system. "Something so revolutionary as harnessing electrical energy required a complete rethinking of everything, top to bottom," Sood said. Not only the physical machines had to change, but the entire process.
That's a crucial insight still relevant to adopting new technology today, he said. The process of factories having to deconstruct how they work in order to capitalize on the use of electricity is a good analogy for what banks must do to effectively apply machine learning to their business models.
Customers browse mobile devices in a Capital One Cafe."It's going to affect every single business process, and it's going to take a fundamental rethinking of how we work today, how we market, how we prequalify customers at Capital One, in order to really take advantage of it," Sood said.
Capital One is up to the task, in Sood's view. He joined the $357 billion-asset bank a few years ago when it bought Monsoon, a mobile design and development company that he co-founded. Going from a startup to a large bank was an adjustment, he said. But he feels Capital One has a startup mentality that traces back to its own beginnings. "In the '80s and '90s, it was using data to offer credit cards to customers that had previously never had access to credit cards," he said. "It was doing marketing and using test-and-learn strategies — which are really statistical learning — in a more manual format."
Machine learning is just a more modern and much faster form of mining data, made possible by the availability of vast amounts of it, Sood said.
When implementing machine learning at Capital One, he opted to create an internal database, then picked one real business application to develop first. Sood's team focused on analyzing clickstream data, which indicates the pages people visit on Capital One's website and what actions they take.
Up to that point, Capital One had been using something like Google Analytics, which had been limiting. "We needed something that was endlessly flexible," allowing teams to set up and run their own queries, he said. "If the mobile apps wanted to track something different than the web apps or the application areas, they could do that. And then, finally, we needed something that for the first time at Capital One would give us data in real time so that we could make decisions — when it came to marketing or anything else — that would happen directly during the same session for the customer."
Now Capital One has a solution that uses machine learning to customize content in real time on the website for every user, based on their behavior during their session. So people looking at a reward card would be served content that is different from people looking at a savings account.
"A more complex use case would involve a credit card application where we detect potential markers for fraud," Sood said. For example, if someone spends too long inputting social security information — something they should know off the top of their head — this would get flagged and dynamically trigger additional questions for the applicant to answer on screen.
Deciding where to apply new technology is key, according to Sood. "My teams call it the 'N = 1 Strategy.' We always look for one use case and try to solve it for that," he said.
Too often companies pursue a solution for every business line across the entire enterprise. Inevitably, this bogs down the process with months of doing due diligence and getting executives to buy in.
"When instead, if you had just put your head down, built a prototype and proved it for one problem, you would have a much higher probability of success," Sood said. "It's not a guarantee by any means, but a much higher probability of success."