Dating apps may get all the press, but NerdWallet has been refining the art of financial matchmaking for more than a decade.
The company provides its members with sound financial advice generated by machine learning algorithms. But as computing has advanced, so has NerdWallet’s ambition.
Now the company is using AI to better match casual visitors to its site — those who haven’t volunteered any personal information — with financial products.
We’re not talking ad stalking like with social media sites. This is predictive modeling, powered by NVIDIA GPUs, that ups the odds of providing valuable recommendations to those who haven’t even asked for them.
The way Michael Tompkins, director of data science at NerdWallet, sees it, if his company wants to provide relevant information to anyone coming to its website, then it needs to make use of all of the data at its fingertips. While NerdWallet has a membership experience, it also draws more than 10 million casual visitors to its site each month.
“We know that people in large numbers are coming, and we know they’re transacting,” said Tompkins. “We need to help them make better decisions.”
Modeling the Unknown
NerdWallet’s machine learning models have traditionally used information provided by members to match them with financial instruments for which they’re most likely to be approved.
The models learn which profile features — credit scores, outstanding balances, credit utilization, etc. — are getting members approved or declined. As models become more familiar with underwriting procedures, they get better and better at matching members with suitable products.
But doing this for non-members presented a more complicated problem: How do you infer without financial information?
The primary resource available to NerdWallet is user behavior, such as what pages people visit, what things they look at and what they click on. By observing these behaviors and matching them against past similar ones, NerdWallet can learn and eventually make better recommendations.
So the company built a model with this in mind and set about collecting “event” information on every digital asset visitors access — from web pages to articles to blogs. With over 100 million unique visitors each year, NerdWallet is collecting billions of events and using that raw data to expand its model.
The company feeds this data into a convolutional neural network. The network’s numerous layers — it can be up to 10 layers deep — are each able to extract factors by creating feature maps of data coming from the preceding layer.
“At the last convolutional layer, we have advanced neural layers and are able to predict or infer what the user might do in their next session,” said Boris Roussev, principal data scientist at NerdWallet. “Or, we can decide based on user behavior what we want to recommend that would be best for this user.”
To accomplish this, NerdWallet relies on the Keras and TensorFlow machine learning frameworks, and its data science team uses NVIDIA V100 Tensor Core GPUs for training. It builds and deploys its models on its machine learning platform that uses Amazon SageMaker.
A big part of the effort involves tuning its neural networks, which requires machine learning models to run hundreds to thousands of times. NVIDIA GPUs shorten that process from days to minutes, enabling NerdWallet to constantly optimize its supporting architecture.
And, as Tompkins points out, every little refinement matters.
“Our particular problem is so tricky to solve that the precision of these models is quite low,” he said. “Getting precision from 25 percent to 27 percent is a huge win.”
Those little gains in precision translate to users having more trust in NerdWallet, which will continuously improve recommendations.
“The grand idea here is that regardless of why you come to NerdWallet, we’re going to provide clarity about financial decisions,” said Tompkins.
The more the company can learn from even the most casual site visit, the more likely it can make useful recommendations quickly. Especially when that behavioral data is combined with the hard data members provide across about a dozen different financial verticals, such as mortgages and insurance.
“If we have this holistic view of the people we do know about,” Tompkins said, “we can use that to make all sorts of recommendations for the person who comes for the first time.”