A 5-Step Framework to Do Machine Learning Right
Machine learning (ML) presents many seductive possibilities to turn data into better business decisions, whether it’s deployed for fraud detection and risk assessment or customer care and predictive analytics. ML applications and platforms are by far the most popular type of AI investment, and according to a 2020 survey fully half of organizations claim to have adopted ML in at least one business function.
Yet in my role leading a machine learning company, I see too many organizations jumping into the deep end of ML without a framework or strategy that will help them make the most of this powerful technology – or even determine whether it’s a good fit for them in the first place. Often there is a sort of blind faith in ML, even if, as a data scientist noted recently, “It’s just math, not magic.”
Since data leaders sometimes can feel on the outside of ML decisions and dynamics – not quite business line executives driving for higher productivity, not quite data scientists working in R or Python every day – I have developed a framework they can use to think about whether ML makes sense and how to bring it into their organization.
The three-pronged ML test
But before even applying an ML implementation framework, I encourage data leaders to take what I call the three-pronged ML test.
Take a moment to consider these questions.
- Is your environment data-rich enough for ML? Companies are always surprised when they discover that what they thought was a data-rich area of their organization simply isn’t, or that a business function hasn’t been collecting data or even thought about that data as a candidate for ML. But here is the kicker, data-rich doesn’t mean “data perfect” – sometimes as little as a thousand records can demonstrate usable patterns.
- Is applying ML operationally relevant? A slightly less polite way to ask this question: Is the problem you’re considering for ML one that anyone in your company cares about? Does it drive business value? I recently had to break it to a data scientist that his business didn’t care that he’d figured out how to reduce the Root Square Mean Error (RSME) of a model by 2%. That business case just wasn’t going to have any impact on the business’s bottom line.
- Is your problem a recurring one? If a data-related problem rarely crops up in a specific part of your organization, it’s probably not a good ML candidate. On the other hand, if you’ve got thousands of IoT devices that are fire-hosing data at dashboards 24x7 but no sense of the business implications, you probably do have a good test case that ML can grind on to find patterns.
A commonsense ML framework
If you objectively determine that you’ve passed the three-prong ML test, you’re ready to move on to the machine learning framework below.
- Target the right type of project
Give some companies an ML hammer, and every part of the business starts to look like a nail. Where a data leader can add value is to pose basic questions: Does ML apply to our organization? And do we have a problem set that’s a good match for it? This means a minimum viable project that won’t break the bank but will serve as a successful proof of concept.
I often see the assumption that you have to launch ML with your biggest, juiciest, meatiest problem, because it seems that this will bring the most value to your organization. The problem is these projects often come with 24-month timelines. When you choose a project that’s too large, you can’t highlight the easy wins that show how much impact ML can have. You want to start with something medium in size, like in the Goldilocks story; not too big, not too small.
- Get stakeholder buy-in
You need stakeholder buy-in from not just your organizational leadership, but from groups who might be affected by an ML project. That’s something else that people often don’t think through completely. ML technology will affect how people are doing their jobs, and likely may make them perform better, but they need some advance notice and dialogue if you want to build their support and trust.
- Don’t fall into the data prerequisite trap
One dynamic I see all too often is that you absolutely need to hire a team of data scientists before you can start anything. Data scientists are amazing, but if you don’t know what business problem you are solving and you think a data scientist is going to figure it out for you, you’ve set yourself up to fail. Just as bad, you may be setting your new data scientists up as data cleansers, in which case they’re likely to leave before a year is out. So before you hire, experiment. There are lots of ways you can implement ML tools that will show easy wins. Then you can make the business case about why you need to attach some rocket fuel to the program.
Another common fallacy is that data leaders feel their data has to be in a perfect state before they can do anything with ML. But data is like art; it’s never going to be perfect. It’s ever evolving based on your company’s needs. If you make data perfection a prerequisite to ML, your organization could fall years behind.
- Get alignment on KPIs
It’s a mistake to assume that your KPIs are about the models themselves. If your business is using rote decision trees, or static if-then statements, but your ML implementation brings them higher revenues, you’ve hit a KPI right there. You’ve added automation to zero. So make sure that the models aren’t the only things you’re looking at measuring. You want to keep bringing the conversation back to, "What is this doing for the business? What value is this creating for the business?"
- Tell a strong story about the data and results
I’ve seen teams fall flat on their faces when they've done something that showed a lot of value but couldn’t make a compelling case. Even if their first foray into ML failed, it probably yielded some learning that's essential for success the next time. So don’t just bring a bunch of stats to leadership. Don’t hedge your results into insignificance. Tie your results back to organizational goals. This is an essential part of being a data leader. Maybe what you’ve done isn’t perfect, but if it lifted your results in a category by 20 percent and nothing you’ve tried has done that, celebrate that story.
A final note to calm the waters
Regardless of industry, you’re likely to hear a chorus of voices in your organization saying that machine learning is coming to take people’s jobs. I think data leaders can play a role in reassuring people that ML is a math-based tool, not a plot to send people out of the company. If you don’t get people thinking that this is going to help them, they’re not going to do it. They may even undermine it.
But ML works. I’ve seen it work countless times. It’s like gravity. When people start using machine learning, they don’t go back, just as people didn’t go back to the abacus once they started using computers. It’s a way to solve problems using the tidal wave of data many organizations have gathered but haven’t learned how to leverage. Finally, machine learning elevates people to focus on strategy while being more productive, which is the white whale that companies continue to strive toward and will continue to do. Just take the time to apply this machine learning framework to make sure it’s the right fit for you.