To Be Successful with Machine Learning, Expect to Learn As You Go
You probably have doubts about your organization's ability to build a successful machine learning practice. Focus on what you can control, and plan to learn as you go.
Got machine learning (ML)? Regardless of whether you do or you don't, you're likely anxious about ML. You probably have doubts about your ability to build a successful machine learning practice. You might worry about recruiting the right expertise, selecting the right technologies, or harnessing both expertise and technology to produce valuable machine-generated insights.
Gartner analyst Svetlana Sicular has some advice for you: don't sweat the big stuff.
If you're smart, diligent, and persistent and you take the long view, you'll acquire the expertise, algorithmic talent, and technology to build a thriving ML practice.
Start out by keeping one precept in mind: you aren't doing ML just to do ML. You're doing ML to improve your business. One of the biggest challenges -- in addition to the paucity of ML-specific skills and the complexity of the extant ML technologies -- is identifying suitable business use-case candidates for ML.
"Data and analytics leaders and their teams excessively concentrate on data collection methods, rather than on understanding and prioritizing business problems to apply machine learning," she writes in a recent Gartner report. "The growing volume of unused data increases their anxiety and they're left with unrealistic hopes that ML will quickly find value in this data."
Anxiety about ML is a consequence of runaway hype about ML. However, the surfeit of hype is merely a distraction and a hindrance; it isn't evidence of some essential problem with machine learning itself.
"Machine learning is in a massive and disruptive evolution that creates anxiety and demands constant learning and adjustments[, but] while tools come and go, people's growing expertise in ML and their understanding of data will pay off over time," she points out.
As you're planning and organizing your ML initiative, there are three questions you should ask yourself.
First, what business problems are good candidates for machine learning? Look for cases in which you can use ML to extend existing analytics investments. "Determine how you will use ML to improve on traditional analytics approaches by surveying the differences between ML and the analytics disciplines you have already mastered," she suggests.
Second, what mix of new or existing human expertise do you need to make your ML effort successful? Sicular suggests building your ML practice by tapping existing statistical and analytics know-how. Promote or recruit statisticians, business analysts, and others with analytics skills whom you can train on ML. "Your task is to find people with math, statistics, or ML experience in your company," she writes, noting that you don't necessarily have to know how to build a predictive model to start doing ML. At first, model building can be outsourced.
Third, how do you start? In building and scaling up a successful ML practice, she argues, you basically have to learn as you go. "You won't have the luxury of learning first, then doing. By the time you acquire knowledge and feel ready to go, that knowledge may be obsolete or irrelevant to your tasks." Sicular's point about outsourcing model building in the beginning is a pragmatic way to both get started and learn on the job.
"The model could be your first ML model on which you are learning, or it could be a packaged application to solve a specific use case -- for example, customer churn or predictive maintenance," Sicular explains. "You can have an external expert ... who can take on the model development task."
Above all, it's important to set expectations: "Set explicit expectations that this is a learning process and mistakes will be made. Perseverance in getting to the right solution in iterations and reproducibility of outputs will bring you stakeholder trust and buy-in for ML."
Stay on Target
It's normal to be anxious about building a successful machine learning practice. If you focus on the stuff you can control, such as identifying and training in-house talent, recruiting new talent, identifying solid, business-centric use cases for ML, and getting started by learning as you go, there's a good chance the rest will take care of itself. There's one more thing, says Sicular: data is the life-blood of ML. If your data management house isn't in order, your machine learning efforts will suffer.
"While tools frequently rotate and change, data remains a cornerstone of the options-based strategy for ML initiatives," she concludes. "The effort of understanding, preparing, and perfecting data for ML works beyond a single project and has a lasting effect by being usable for building many models. [Your ability] to understand and manage the data lessens the ML anxiety."