Machine Learning Times
Machine Learning Times
EXCLUSIVE HIGHLIGHTS
SHARE THIS:

4 weeks ago
Why Operationalizing Machine Learning Requires a Shrewd Business Perspective

 

Originally published in Analytics Magazine

For a rocket scientist, the math isn’t the hardest part. What’s hard is being so often misunderstood.

The same goes for data scientists, who time and again lack the support needed to successfully launch the fruits of their brilliant labor into action. These math heads have got to integrate into the organization as a whole, lest they vanish into the obscurities of their analysis. Their isolation is an enemy to their usefulness.

After all, the most wicked and pervasive pitfall of predictive analytics is organizational in nature, not technical: Predictive models often fail to launch. They’re never deployed to drive decisions. This is ultimately a management error. We must pursue the business of machine learning only so that it delivers the business value of machine learning.

Here’s how the delivery of that value works. Operational decisions need prediction. Prediction requires machine learning. And machine learning requires data. Reversing that into sequential order, we have data, we give it to machine learning, it makes models that predict, and we use the predictions to more effectively drive millions of operational decisions. With this in place, we improve all the large-scale organizational functions that make the world go ’round.

data→ machine learning→ model→ predictions→ operational decisions

The part on the far right, the predictive model’s intended operationalization – aka, its deployment or implementation – guides the entire machine learning project. How the model will ultimately be used is the carrot at the end of the stick. It keeps the project moving in the right direction from the get-go.

In a nutshell, we must steer the project toward the intersection of two sets:

On the top, the range of conceivable prediction goals that would be valuable for digital decisioning is limited only by the imagination of your operations and marketing staff. Prediction goals could be any of a range of behaviors and outcomes, such as which customer will buy or which transaction is likely to be fraudulent. The problem is, some of these ideas aren’t doable from an analytical standpoint.

On the bottom, there are many things that could be analytically predicted. So long as there are enough examples in the data, machine learning can generate a predictive model that puts future odds on individual cases. The problem is, many such ideas that sound appealing in the lab would never actually be used, never operationally deployed. Your team may create an elegant, effective predictive model, but that doesn’t mean the business is ready to act upon it to drive decisions. All too often, lack of management buy-in or unforeseen business constraints preclude model deployment.

A shrewd business perspective steers clear of this pitfall and navigates to a viable prediction goal within the intersection of these two sets – one that’s both achievable and useable. The analytics and number crunching alone do not determine what to actually do with a model’s predictions – only business acumen can dictate how to best deploy a model. For example, for targeting a marketing campaign, a predictive model provides a range of options and trade-offs. Such a model could empower you to market to a group of, say, 100,000 customers with a four times greater likelihood of making a purchase – or, alternatively, to a group twice as large but with only three times the likelihood. You’ve got to choose between these options based on business considerations such as profit and market penetration.

That’s where the book for which this article originally served as a foreword comes in: Digital Decisioning: Using Decision Management to Deliver Business Impact from AI. Author James Taylor is the longstanding thought leader in machine learning operationalization – the deployment of predictive models to drive business decisions. He and his firm bridge what is often a wide gulf between business leaders on one side and hands-on analytics practitioners on the other.  For many years, James has also brought his insight to Predictive Analytics World, the conference series I founded, where he serves as the co-chair of the operationalization track on machine learning deployment, as well as the instructor of the event’s training workshop on this topic.

This comprehensive book guides you to leverage the potential of machine learning. It delivers the business-level finesse needed to ensure predictive models are operationalization-ready. It lays the groundwork and sets the standard. It’s a great place to start… and to finish.

About the Author

Eric Siegel, Ph.D., founder of the Predictive Analytics World and Deep Learning World conference series and executive editor of The Machine Learning Times, makes the how and why of predictive analytics (aka machine learning) understandable and captivating. He is the author of the award-winning book Predictive Analytics: The Power to Predict Who Will Click, Buy, Lie, or Die, the host of The Dr. Data Show web series, a former Columbia University professor, and a renowned speaker, educator, and leader in the field. Follow him at @predictanalytic.

 

Pin It on Pinterest

Share This