The fundamentals of supervised and unsupervised learning

Extracting value out of data is a universal challenge for insurance professionals. Colby Tunick, CEO and co-founder of ReFocus AI, has set out to help insurance brokers and carriers use their data to improve their customer retention by proactively identifying which policyholders are unlikely to renew and preemptively coordinate renewal outreach.

‍

“As you can imagine, most insurance carriers have mountains of data, and sometimes they’re not quite sure how to use it,” says Tunick. “That’s where we come in.” Tunick further clarified that there are other problems that could be solved using existing data, but retention is their niche.

ReFocus AI, a BrokerTech Ventures (BTV) 2023 accelerator alum, helps insurance agents, brokers and carriers reduce churn rates by using existing policyholder data to detect customers who are likely to leave. Finding these customers before they leave enables businesses to reach out proactively and improve profitability. 

Here, Tunick answers some fundamental questions about how insurance professionals can embrace AI:

Why is everyone talking about AI?

“Everyone’s talking about AI because it’s in the media. But when you look into the history of AI, it’s been around in a similar form since the 1960s. The same algorithms and techniques used to extract value out of datasets have been around 80 years. But AI has been limited by the availability of data and computing power. So, in the 1960s, the standard was maybe one megabyte of data. Now, we’re generating petabytes of data every minute. We also have cloud computing, so we’re able to harness almost unlimited computing to extract patterns and trends. In short, the rise of large available datasets and computing power is what is driving innovation and all of the buzz in the news. Now, we’re able to make use of the data.”

What is the biggest problem with machine learning? 

“There’s a common refrain that more data equals better insights, but that’s counterintuitive because there are so many different ways to do machine learning and so many types of datasets you can use. You have to know the problem you are trying to solve in order to select the right data, the right method and the right way to optimize it for results. 

‍

A lot of companies are migrating to the cloud and they’re saying, ‘Tell me what we can do with this.’ Then, they’re getting results that they don’t know what to do with or use to drive value. 

‍

For example, maybe you learn that a percentage of your prospective customers prefer to drive a blue car. That’s a statistically valid insight, but how is that going to drive your customer acquisition strategy forward? Hint: it probably won’t. That’s the problem with machine learning. You have to have a business outcome in mind in order to select the right ingredients to make the cake come out right.”

What is supervised vs. unsupervised learning? 

“Supervised learning is a field of machine learning where you have labels and previously labeled data. You’re able to then show examples of that labeled data to a machine-learning algorithm, and it is able to detect those labels when it sees them. Supervised learning happens when the person tells the machine what the patterns are and labels them. 

‍

With unsupervised learning, there is no human telling the machine what it should be doing. Therefore, the algorithm must learn to self-discover any naturally occurring pattern in the dataset. Unsupervised learning is good at finding patterns in large groups of datasets that are similar to one another. Unsupervised learning helps overcome the limitations of supervised learning, but it also tends to be less precise and more difficult to do.”

To learn more about the different types of learning, along with helpful visuals, Tunick recommends an article written by his colleague and cofounder of ReFocus AI, Nisar Hundewale, Ph.D., “How Machines Go From Dumb to Smart by Learning.”

‍

What are the insurance applications of supervised vs. unsupervised AI models? 

“Most companies start with supervised learning. These models are good at providing a binary decision or a categorical classification. If your business problem requires a yes or no answer, it’s supervised learning. Or, if your business problem requires one of five different outcomes, or you’re trying to assign an outcome one of five different ways, that is a great example of supervised learning. For example, is a customer likely to buy? The answer is yes or no. Is a customer who bought two products in the past likely to buy a third, fourth or fifth? The answer is categorical. 

‍

Unsupervised learning is popular for extracting value from video, audio and large amounts of text. For example, you have people take a photo of their roof, and the AI will determine how old the roof is based on weathering, materials and other factors.

‍

Other common applications for AI in insurance include identifying fraudulent versus real claims, or assigning the right amount you should offer to settle a claim. These are well-defined problems a lot of companies are doing.”

‍

What are the challenges of each model?

“With supervised learning, you need a lot of labeled data, so it’s time intensive. Unsupervised learning doesn’t require all the manual data labeling, but it’s not as sophisticated, so it may not lead to a usable outcome (also called convergence).”

‍

How do you use AI to improve retention?  

“We use structured and unstructured data in a supervised model to predict someone’s likelihood to not renew their current policy. Our client provides us with labeled data (e.g., known historical outcomes) from their CRM of current customers who have renewed, along with current or past customers who have canceled. We then go back 5-20 years and the model is able to tell us — for upcoming customers it hasn’t seen before — are they are similar to the people who did renew or similar to the people who did not? So now, our client can focus their attention on the customers who might not renew and automate the renewal process for the rest.” 

‍

What are the ethics around supervised and unsupervised AI learning?

“First, it’s important to remember that all data is biased. For example, let’s say our customer sells 80 percent of their business to car washes and 20 percent to restaurants. The data will be biased towards car washes because it has seen more of that data. It’s not good or bad, it’s just biased. For example, in machine learning, you can say the model is more heavily weighted one way or the other based on the distribution of the data (e.g., car washes or restaurants). 

‍

Second, it’s important to understand data that may be a proxy for information that is unallowable. For example, when selling an insurance policy, using a credit score is unallowable. But sometimes other data, like a zip code, may be a stand-in or proxy for the unallowable information.  

‍

Third, we have to understand the inherent limits in the data we’re using. Sometimes we want the output of the machine learning model to be the end-all be-all, but perhaps there is other, nuanced information not in the dataset, and therefore what we produce with the data isn’t actually a solution to the business problem. Just because you build it doesn’t mean it will come or it will be useful.”

‍

The key is to focus your efforts

When it comes to choosing between supervised and unsupervised learning, the one you use depends on the outcome you want. “Think of supervised and unsupervised learning the same way you’d think about the different types of wrenches in your toolbox,” says Tunick. “One isn’t better than the other. What matters more than understanding which model to use is being outcome driven and applying the right technique to get there.”  

‍