putilov_denis - stock.adobe.com
At Capital One, machine learning has become a key part of its business, with the financial services company adopting a standardized process for developing models and sponsoring research to help define its strategy.
The McLean, Va.-based financial services company deploys ML in several use cases, building upon its cloud-based data ecosystem. Support for ML goes all the way up to the C-suite: Rich Fairbank, CEO at Capital One, mentioned ML seven times during an analysts' call last year and cited the use of ML for monitoring the economic environment. CapitalOne now pursues the emerging practice of ML operations (MLOps), essentially, DevOps for ML, to further institutionalize the technology.
"Our AI ML capabilities are absolutely central to how we build our products and, even more than that, they're very important to the way we actually run our company," said Zachary Hanif, vice president and head of machine learning model and platforms at Capital One. "We have been, over multiple years, leveraging machine learning capabilities across the business in diverse ways."
Those ways include using ML to bolster fraud detection, deliver more personalized customer experiences and improve business planning. As for the latter, "we're making sure that we have a better understanding of emerging market conditions and our place within the larger economy," Hanif said.
Machine learning in banking
Capital One is not alone in pursuit of ML. Large banks are leading the way, creating their own infrastructure for spinning up applications.
But smaller financial institutions are also looking to tap ML, using third-party platforms and services rather than building in-house capabilities.
"I think it's important to all banks right now," said Joe Davey, a partner in the technology practice at West Monroe, a consulting firm with headquarters in Chicago. "Banks, overall, have been trying to leverage technology to reduce their efficiency ratios," he said, referring to the ratio between operating expenses and income. "[ML] is just one more piece of the automation puzzle."
MLOps at a glance
The MLOps discipline defines processes for governing the ML lifecycle, from identifying business objectives to developing, training and deploying models.
MLOps uses DevOps approaches, but enlarges the scope of the latter methodology to take on complicated model creation and delivery tasks. MLOps also differs from DevOps in that it aims to promote collaboration among data scientists, operations teams and business professionals.
"Clear standards and automated processes" help break down organizational silos, according to the Capital One-sponsored HBR report. Technical contracts play a role here, providing well-defined connections between systems, Hanif noted. Those connections enforce particular standards such as gRPC, which lets developers specify a service endpoint and the data format it will accept. This approach ensures clients and servers comply with a given specification.
Building an ML platform
Capital One's current ML initiative stems from a decade-long technology transformation -- a program that included re-architecting its data environment.
The resulting cloud-based platforms -- Capital One uses Snowflake's data warehousing and engineering platform, for example -- provide the baseline infrastructure on which developers can build and deploy models.
"Infrastructure allows your teams to focus on the problem that they have at hand without thinking about all the necessary components that are required to support solving that," Hanif said. "Developers spend more of their time focusing on the material that matters most to the business problem."
He said the platform approach also advances the computer science concept of accessibility, which aims to make data and applications intelligible and available to users and developers.
"Accessibility is incredibly important," Hanif said. "If you cannot make a piece of software accessible to your users -- meaning that they're able to understand it, can think about how they can apply it, and can see a use for it inside of their environment -- it has, in essence, failed to deliver on its promises and potential."
Platforms become important when organizations seek to expand AI and ML beyond early experiments and pilots. In the pharmaceutical industry, Eli Lilly created an Enterprise Data Program and centralized analytics platforms to help scale AI across the company.
Ramping up MLOps: Challenges and benefits
Scale is a matter of method as well as technology. In that vein, MLOps provides an approach for running an enterprise-wide ML program. Hanif said Capital One has "fully adopted MLOps processes" and ranks among the early adopters, particularly within the financial services sector.
"We see MLOps as the foundational framework to be able to set up teams for success inside of machine learning, deploy their capabilities at scale, and to ensure that we're able to create an end-to-end environment," Hanif said. The goal: providing a consistent environment for designing, deploying and managing ML models, repeatedly and in larger volumes.
Zachary HanifVice president and head of machine learning model and platforms at Capital One
Unruly data is one barrier in the path of MLOps and at-scale ML. Organizations may have data stored in a variety of places, making it hard to discover, Hanif said. "The first challenge you must always engage with is data," he noted. A second challenge, Hanif said, is establishing an ML workflow that development teams can follow.
Organizations that overcome those obstacles could potentially see a significant increase in ML efficiency. A Harvard Business Review Analytic Services whitepaper, citing data from Cnvrg.io, noted that earlier adopters of MLOps reported as much as a 10-fold increase in productivity and 5-times faster model training.
That report, which Capital One sponsored, also suggested most ML models exist outside of a structured process for managing them. The report cited IDC research that contends 90% of ML models are not deployed into production.
Such models may actually find use in organizations and affect business decisions. But they aren't deployed inside of a standard release pipeline with large-scale automated testing and validation monitoring, Hanif noted.
"You have data scientists who are developing hundreds or thousands of ML models that never really quite see the light of day," he said. "They exist in a kind of shadow state."
That said, Hanif said he believes more companies are now exploring MLOps to create a well-governed framework for the ML lifecycle.
This structured way of managing ML arrives as more financial institutions recognize the technology's potential.
"Banks are starting to understand those use cases better than they did a few years ago," West Monroe's Davey said.
He pointed to anomaly detection and credit risk as typical financial services applications, noting that all large banks and many, if not most, midsize banks pursue those applications. Document processing and onboarding, meanwhile, are emerging use cases on the operational side, he added.
ML hiring continues
Despite economic headwinds, Capital One's tech organization continues recruiting for talent, including engineering roles within the cloud, data, ML, cybersecurity and product management fields.
Some roles have seen cutbacks: Capital One earlier this month eliminated positions in agile development, according to several news outlets. The reorganization, however, reflects how agile processes have become part of the company's core engineering practices and doesn't affect areas such as ML, according to the company.
Investing in learning
Another aspect of Capital One's ML strategy is sponsored research, the HBR Analytic Services whitepaper providing one example. That report, released in October 2022, built a case for the MLOps practices Capital One is following: "Companies without mature MLOps programs could find their competitors outpacing them in using ML," the whitepaper stated.
Capital One, also last year, commissioned a Forrester Research report on ML challenges. For that research project, Forrester surveyed 150 data management decision-makers in North America. The report highlighted anomaly detection as the top ML use case and pointed to the importance of partnering with third parties to advance enterprise ML strategies.
The investment in research informs Capital One's ML methods and technology platforms.
"We're making sure and we're developing a whole body of learnings to ensure that we're leveraging best practices," Hanif said.