One of the earliest distinctions you may encounter in machine learning is that between “supervised” and “unsupervised” machine learning. The distinctions between these two types of learning are important as they largely define the kinds of real world problems each is best suited to solve.
Let’s start with supervised machine learning. A supervised machine learning problem is one where the computer is trained to predict future outcomes by studying the relationships between prior inputs and prior outputs. In this kind of problem, we know what specific question we are trying to answer- what outcome we’re trying to forecast or predict- and we are asking the computer to learn enough about our data to start answering those questions for us in the future.
For example, a marketer may want to predict which coupon or offer they should present to a prospective customer to inspire a signup. The marketer knows their desired outcome: a sign up. They are relying on the computer to ingest everything it knows about this prospective customer, as well as everything it knows about prior prospective customers that either did or did not complete the signup, to offer a prediction about which final offer will likely yield the desired outcome.
In this kind of problem, you can imagine that we have a historic dataset of all prior customer engagements, where each row represents a different customer, and each column represents a different characteristic about that customer, or their interactions with our company. Column A may be their geographic location; Column B may be the days it’s been since they last visited our website; Column C may be what discounts, offers, or coupons they were offered, etc. Finally, the last column would be our target: did this customer signup?
Using all of this historic data, our computer is studying the subtle relationships that may exist among all of these various characteristics and activities among old customers, in service of developing a model that can predict a future signup based upon those characteristics and activities of new customers.
This is a supervised machine learning problem because we have supervised, or directed, the computer to look for patterns to build a model that predicts a particular outcome, and we have provided it with all the data it should need to find those patterns and to make those predictions.
So what about unsupervised machine learning? By contrast, an unsupervised machine learning problem is one in which we have no particular desired outcome in mind at the start. Instead, we are asking the computer to consume all of our data to identify patterns that would suggest potential relationships or similarities within it that may be of interest to us.
Consider again our marketer from the prior example. Perhaps they wish to leverage all of their prospective customer data to create personas that they can use to better think about the different types of people that engage with their marketing materials. They are relying on the computer to notice these clusters and to automatically group data points together that share similar characteristics. Now, perhaps, the marketer can use these generated groups to create data-driven marketing personas that can inform how they segment and target future outreach to different types of prospective customers.
This is an unsupervised machine learning problem because we are giving the computer free reign to surface relationships within the data, without a predetermined outcome in mind.
It’s important to note that not all instances of data clustering are unsupervised learning problems, however. If, for instance, the marketer began the project by defining the 7 personas that all customers must fall into, and then used machine learning to auto-classify each new prospective customer as belonging to one of those seven groups based on how previous customers had been classified, this would be the work of supervised machine learning. That’s because there are a defined set of outcomes that the computer is taught at the outset, based on prior data. Unsupervised machine learning, by contrast, is given no such direction and merely organizes the data as it identifies similarities.
As you can tell by their differences and the above examples, whether or not a supervised or unsupervised approach to machine learning is appropriate in a given context has a lot to do with what the intended outcome is, as well as the available data. In situations where there is a clear outcome, answer, or solution and you hope to use machine learning to arrive at it, a supervised approach is likely best. But you’ll need sufficient past data to teach the computer the difference between “right” and “wrong” answers or outcomes. In situations that are more exploratory- where there is not necessarily a “right” or “wrong” answer or outcome- an unsupervised approach can be used to learn more about your data.
Of course, in many practical applications of machine learning you may encounter circumstances where a combination of approaches are utilized. Let’s return one last time to the marketer from our examples. A smart first step may be to use an unsupervised learning approach to cluster the available data of prior users and prospective users. Then, our marketer could study the groups that are surfaced as part of the process of creating personas. Then, a supervised machine learning model could be used to auto-assign incoming prospective customers to one such persona which could dictate the different marketing content and journey that individual is exposed to.
Learn more about applications of machine learning, and other ML 101 topics on the Telepath blog, here.