How Machines Learn (And You Win)

"Machine learning.” You’ve heard the term, and you probably nod in agreement when someone tells you how important it is. But secretly you may not be sure what it is or how it works. Ask your data scientists to explain, and you may get lost in a sea of specialist talk about forks, leaf nodes, split points, and recursions. Forget all that. The only thing you need to know is that machine learning applies statistical models to the data you have in order to make smart predictions about data you don’t have.

Those predictions can help you find signals in the noise and extract value from all the data you’re collecting. The advantage of—indeed, the imperative for—using machine learning is its speed and brute force. It can mine vast swaths of data in seconds or minutes, find patterns, and make predictions in ways that no human analyst could begin to emulate. Machine learning is, among other things, helping companies to detect that patients will have seizures long before they actually occur.

Best of all, no matter how much work you throw at it, the student never gets exhausted or bored. Here’s a look at how it works.

Example: How a Cable Company Learns Which Customers Might Cancel Service

A cable company’s defectors have started to outnumber new customers. The company wants to launch discount offers to prevent more cancellations, but it doesn’t want to offer discounts to satisfied customers who are willing to pay their current rate. The company has no explicit data on “potential cord cutters,” so it will use its customer database and machine learning to spot likely defectors and target only them with its discount offers. We’ll follow the machine’s learning process as it tries to sort two customers—Karl and Cathy—using one typical ML exercise: a decision tree.

1. Process: Decide What You Want to Know

A machine can’t choose to learn something on its own (yet). Instead, data scientists tell it what they want it to learn.

2. Build a Model

To learn, the machine must access rich customer data: demographics, purchasing behavior, other services customers pay for, whether they’ve called to complain or cancel a service, and much more. Using these variables, data scientists build a statistical model that determines how predictive each variable is in terms of the answer the machine is trying to learn.

3. Accumulate Answers

For each question, the probability is refined in one direction or the other on the basis of answers up to that point. The machine looks for combinations of attributes that create a high level of certainty about the answer it’s seeking.

4. Make a Decision

Eventually, the probability is heavily weighted one way or the other. That may come after just a few questions or dozens of them. The data scientists decide how confident they want the machine to be in its answer. They may say that once it’s 95% confident, it can stop.

5. Iterate

Now the computer can carry out the same evaluation thousands or even millions of times on other entries in its data set, confidently sorting each entry into group A or group B. It has learned.