A New Frontier: How Machine Learning Explanations Will Improve Your Business

November 18, 2016
· 2 min read

by Daniel Becker
Data Scientist | Technical Product Director at DataRobot

Everyone expects machine learning models to make accurate predictions. At DataRobot, we’ve helped our users develop models to make a wide range of predictions, such as:

  • Which customers will cancel their service?
  • Which baseball players will become superstars?

What fewer people realize is that machine learning models can not only answer questions but can simultaneously deliver powerful insights. Our users have been surprised (and impressed) when we help answer questions like:

  • What are the primary drivers of service cancellations?
  • Why will a certain player become a superstar?

Reason Codes, introduced in the latest update to the DataRobot platform, provide clear explanations for every prediction in a way that most data scientists never imagined possible. The result is greater than the sum of its parts.


To see the value of prediction-level insights, let’s go back to the predictive questions described above.


Which customers will cancel their service?

Many businesses ask this question and then assign at-risk customers to a department tasked with minimizing churn. The customer outreach team contacts all potential churn customers with a generic message, hoping it will address each customer’s underlying reason for potential cancellation.


With reason codes, the outreach team can proactively identify and address each customer’s concern. A customer worried about cost might receive a pricing promotion. A customer upset about a service glitch may be contacted directly by technical support.

With reason codes, “shot-in-the-dark” contact can be replaced with a targeted and effective retention plan.



Which baseball players will become superstars?

The success of the 2003 Oakland Athletics, chronicled in the book and movie “Moneyball,” has made data science increasingly prevalent in professional sports. Hoping to leverage those same successes, teams have come to rely heavily on predictive models to identify upcoming superstars. Reason codes open up a whole new range of applications with these models, extending the forecasting–and hopefully the season.

With a corral of young players, how do you identify those that are good enough to make the majors? You can build a predictive model that says, yes or no, with some percentage of likelihood for each player. But what if you have invested in a star whose stock is plummeting? Reason codes can help identify the reason for that drop—pitching too much given injury history? From the answers you can adjust the environment and potentially revive your season.

One of the biggest gripes that users have with machine learning systems is the lack of visibility into why decisions have been made. With DataRobot reason codes, you can now extract insight from the models and make decisions based on this deeper level of visibility. If you want to see how reason codes work, you can request a demo to see DataRobot in action.


New Call-to-action

About the author
Dan Becker
Dan Becker

VP of Decision Intelligence, DataRobot

Dan Becker is the VP of Decision Intelligence for DataRobot. Dan finished in 2nd place (out of 1350+ teams) in a machine learning competition with a $3Million grand prize. He subsequently led AI consulting projects for 6 companies in the Fortune 100, and over 500,000 people have taken his applied AI courses on Kaggle. Dan has contributed to leading open source AI tools including TensorFlow and Keras. More recently, he co-authored the textbook Automated Machine Learning for Business. Dan has a PhD in Economics.

Meet Dan Becker
  • Listen to the blog
  • Share this post
    Subscribe to DataRobot Blog
    Newsletter Subscription
    Subscribe to our Blog