Using Predictive to Be Proactive: Machine Learning 101

Machine learning isn’t a new concept. The term was first coined in 1959, when computer gaming pioneer Arthur Samuel described it as a method for giving “computers the ability to learn without being explicitly programmed.” This is still a fair description today. What has changed in the almost six decades since the term was born, though, is how effectively machine learning can be used by companies when engaging with customers. Predictive analytics is a powerful subfield of machine learning, and businesses that can accurately forecast the behavior of their contacts are in prime position to acquire, retain, and evangelize people.

There are three types of machine learning: supervised, unsupervised, and reinforcement. For marketers, the two main flavors tend to be supervised and unsupervised. Unsupervised machine learning occurs when a system can freely determine patterns in data, with no specific target or goal informing the hunt for patterns. (In fact, it’s perhaps best thought of not as a ‘hunt’ and more as a ‘watch, wait, and see.’) This technique is probably most famous for fueling recommendation engines, since it can help customers discover brand-new products and can help businesses unlock non-obvious (and profitable) groupings.

This post will focus on supervised machine learning, a technique that applies to systems that utilize predictive analytics. But first, let’s make sure we can agree on what the term “predictive” actually means. The marketing punchline is well worn by this point: the demand generation, content, and social teams, all huddling over the glowing crystal ball so they can dispatch the perfect asset to the perfect audience at the perfect time. While this (however silly) is what we’d love, predictive analytics is not about being perfect. It’s about being probable, or having a ‘better than luck’ chance of correctly making a resonant, personalized journey for contacts. Also, predictive is not just about reading the future. It’s also about extrapolating the present.

As a subset of supervised machine learning, predictive lets companies use the data they already have to [a] guess other or unknown qualities about contacts that are probably true (present) and [b] guess how people will behave (future). This helps companies market proactively instead of reactively, which can increase people’s interest from the get-go and boost your odds at converting. In other words, modern segmentation relies on predictive analytics, and prediction relies on supervised machine learning.

Unlike unsupervised, supervised machine is searching for a specific answer—often looking to fill in a certain data point. Also unlike unsupervised, supervised machine learning is expected to yield an answer that is logical according to what the system already knows. That’s because in supervised, a business constructs a model with the information it has (called build data or training data), and that model matches people against patterns. So, if a part of a customer’s digital story or demographic background is missing, the model can tell the business what data point probably fills that blank (or will fill it in the future).

There are two types of supervised machine learning: classification and regression. Classification occurs when the target answer is a descriptor or category; when the machine has learned enough about the customer (and the community they’re part of) to classify them as Year-Round Shopper, for example, or Holiday Vacationer or Credit Card Maxer-Outer. Regression occurs when the target answer is a number value; when the question isn’t “What kind of shopper is this person?” but rather, “In how many different months did they buy something?”

Let’s say, for example, a clothing retailer wants to automate the first message it sends to people after they make their first purchase, and it wants to base that messaging on whether customers are likely to become Big Spenders or Little Spenders. (Not that machine learning is limited to binary outcomes. Anything but! Results can theoretically be infinite.) The retailer could make an algorithm based on several factors: age, location, money spent on that first purchase, if the person added a recommended item to the cart, and the amount of days the cart was abandoned. If the database knows the answers to all these questions, the model can match each customer to a pattern of behavior that signals ‘Big Spender or ‘Little Spender.’

This example would qualify as the classification subtype of supervised machine learning, since the output was descriptive, not numerical. However, if the retailer weren’t segmenting by category but by number value—as in, not predicting if people become Big or Little Spenders, but estimating the lifetime value of purchases they’ll make on the site—it would be a regression use case. It’s not uncommon for classification and regression to be two sides of the same coin, or different angles of approaching the same core business question.

And again, don’t let the word “predictive” enamor you to the future. The retailer in this example might already know someone’s a Big Spender: they live in an alpha city, spent more than $150 on their first purchase, bought at least one recommended item, and checked out same day. How old are they? Even if this data point is officially unknown, the algorithm could predict age if the other data aligns with a known pattern.

With the rise of automation and big data, machine learning plays a crucial role in marketing today. This post, of course, has only scratched the surface. But I hope it’s made clearer the connection between segmentation and predictive. Good marketing isn’t necessarily perfect, but it is proactive—and supervised machine learning helps that happen.

If you need support with implementing automation, please contact us!  We love helping companies transform their marketing automation.