Global Data Science Forum

ML Operationalized: An Enterprise Story

By William Roberts posted Thu March 15, 2018 05:39 PM


Machine learning is a technique, and data science is the practice. That’s a familiar narrative, and it helps ground leaders new to the field in an understanding of what they’re trying to execute. The view from the field nowadays, though, is that the greatest challenge facing the enterprise comes less from enabling the practice, and more in how enterprises operationalize it. What do I mean by this?

In theory, machine learning is a technique for cleaning data, choosing algorithms, and training models. The output of those models is a prediction, and those predictions are valuable in improving business processes across industries. Leaders knows they need to implement machine learning. They even know they have the data — 80% of all of which is estimated to sit behind corporate firewalls. How does reality contrast to the theory?

With existing IT architectures, data isn’t typically clean and ready to be consumed. Leaders on analytics teams may or may not be enabled to run the model training their enterprises need, validate them, test them, and then deploy them. Finally, there’s the act of managing those models that whole business groups have worked tirelessly to deploy: As models start to lose accuracy, teams can run into roadblocks with internal business processes as they try to revise or replace the models. That reality has refined IBM’s perspective of the operationalization of machine learning into what really has become an enterprise story.

In an enterprise, simply adding more data scientists isn’t a cure for all corporate woes. That remedy requires thoughtful team structuring, and smart tooling in harmony with each other. Data scientists belong in a broader team with analysts, data engineers, and production engineers. Where data scientists might lack industry knowledge, data analysts inform — based on years of experience and precedent. Where analysts might not know why a random forest is more appropriate than a logistic regression, data scientists can direct.

Scaling up the machine learning ladder takes a full spectrum of skills, but there’s also a scarcity of data scientists available to help. A big part of the solution is making sure that the data scientists you do have on staff can be as efficient as possible. That means using enterprise-grade versions of the open-source libraries they love — and ensuring they’re integrated. Whether clicking or coding, EDA, training, validation, and testing all need a common platform for execution.

If one thing is clear, it’s that operationalizing machine learning won’t be a one-step process by any means. It’ll take planning, patience, and craft, but the effort is worth it, as IBM has demonstrated again and again with our clients. When enterprises embrace ML, they can push their predictive powers in concert with prescription. A company can’t prescribe how many snowboards to inventory if they can’t predict when there will be snow to snowboard on.

There’s harmony in leveraging descriptive, predictive, and prescriptive abilities, and IBM has a lot to say on the subject. I talked above about the struggles, and the outline for enabling a solution. We want to tell you more at THINK 2018 about how IBM enables enterprise-grade machine learning to put your company on the path to AI. Register today.