How to Improve the ROI of Your ML Deployment

Video produced by Steve Nathans-Kelly

At Data Summit Connect 2021, Pythian SVP Analytics Lynda Partner described how to deploy machine learning in a way that will keep your CFO happy—regardless of your current stage of ML maturity.

For companies in the early stages of ML maturity, Partner offered four pieces of advice to attendees.

"One is spend more time on your space for use case selection. I can't overstate this enough. And to do that, it's a team sport. You need IT. You need data science. You need the business. Invest in a cloud data platform—this is the thing that is going to organize your data for you and speed up everything once you have selected your use case. You need IT for this, the business can't do this, the data people alone can't do this."

Once you have selected your use case, she explained, you are going to want to start a data governance program and you will want to make sure that your use of data is consistent across all of the use cases. "And, for that, you need IT and business. And you need to educate more people about machine learning. And this is all in the interest of reducing the communications gap and understanding good use cases and bad use cases,  understanding what machine learning means and how it's going to change the busioness means a lot of talking; that's how you get started."

And if you've already done the above and you want to take the next steps, Partner offered three pieces of advice: First, invest in an integrated development environment. "You need IT  for that, but it is going to take you from a really long time to get through this process to a much shorter time and it's going to eliminate a whole bunch of costly mistakes through the process."

In addition, she said, you need to invest in MLOps skills, tools, and processes. "We have seen so many of these just get stalled at the point where they have their model—they just don't know what to do with the output. So the ability to actually take that model and deploy it production and run it at scale and have it operate without you having to worry about it is super important."

You don't want to get to that point and then have everything grind to a halt, she said. "You need IT for that. And then start thinking about model management models." It is important to ensure that these things continue to be tuned, optimized, and tweaked. "And for that, you're going to need your data science people. No one group in a company can ever say that they can improve the ROI and make things a success. It has to be a collaboration of all of these people."