Machine learning is a buzzword you may have heard around your office or on the news. It’s an emerging technical trend that can seem intimidating, for good reason. The most obvious being the specialized expertise required to develop and apply the algorithms that make it possible for machines to “learn.” However, as data science skill sets become more available and toolsets increasingly easy to use, organizations have realized the artful application of machine learning requires more than just technical expertise.
Just as with the previous generation of advanced analytics solutions, value comes from putting found insights into action. So whether you’re a beginner just starting out or are looking to expand a nascent machine learning program, these 10 (nontechnical) best practices can help ensure success.
1. Educate the business on concepts, not theorems
Making the case for time and funding of big machine learning projects requires nontechnical executives and businesspeople to broadly understand what machine learning can and can’t do. But that doesn’t mean they need to understand the nitty-gritty details of the latest neural network.
A story that demonstrates how machine learning can be applied to your business will garner more engagement than complicated algorithmic charts and discussions of p-values. Share examples of problems machine learning can solve, while including case studies from other industries and like companies. If you must explain the method itself, think analogies – not engineering diagrams.
2. Make machine learning part of the discovery process
When getting started with machine learning, it’s important to understand what your needs are and what machine learning can do. In some cases, machine learning may identify areas for further study and consideration. In others, machine learning algorithms themselves might be integrated into operational systems to automate key decision points or processing pathways in real time. Business and policy decisions must consider this when deciding if and how to put found insights into action.
3. Avoid black box exercises
Yes, machine learning methods seem obscure and are, in fact, often inscrutable. But machine learning is not a black box activity in which results can be blindly applied without due consideration.
Humans, above and beyond the data scientist programming the algorithm, are required to consider and answer questions such as:
- What are we trying to predict?
- What is the best likely input into the process?
- Are results in line with expectations?
- Are there exceptions to be addressed? What are the implications if they’re not?
- What is the proper response? How can (and should) results be applied?
The Machine Learning Primer: A SAS Best Practices e-book
Download the comprehensive machine learning e-book, which includes machine learning techniques and a checklist for getting started.
Free machine learning e-book
4. Apply appropriate scientific rigor
While extremely powerful, machine learning isn’t a magical, self-correcting analytical panacea. In practice, developing a machine learning application is an experimental and iterative process. This is true even when using well-known algorithms. In every case, the algorithm must be trained on and tuned for the business context and data at hand.
Whether you’re a machine learning beginner or an experienced practitioner, a healthy dose of skepticism and solid statistical reasoning, scientific and data analysis is critical.
5. Make it just complicated enough
Generally speaking, more data wins. But do more features (aka attributes or data points) also equate to better outcomes? Not always. The catch is that larger data sets beget larger variations in data and a larger potential for predictive error.
In many cases, weaker models with more data do better than more complicated models with less – even if the data is dirtier. This is the basis of ensemble modeling techniques that use the “power of the collective” to predict results.
Deploying machine learning artfully is a balancing act. One in which the incremental predictive value of complexity must be weighed against interpretability, ease of use and applicability.
6. Actively engage business users in validation
Data scientists must apply due diligence as the model is created and tuned. To do so, they must effectively communicate and collaborate with data and business domain experts to validate and vet the model.
This is critical to ensure the team has, among other things:
- Validated that all the options have been considered.
- Accounted for potential bias.
- Identified the impact and implications of putting found insight into action.
Whether you’re a machine learning beginner or an experienced practitioner, a healthy dose of skepticism and solid statistical reasoning, scientific and data analysis is critical. Kimberly Nevala Director of Business Strategies SAS Best Practices
7. Employ data storytelling
The insight created through machine learning may appear as obscure or unexpected as the algorithms themselves. Particularly when those insights challenge existing business wisdom or affect standard operating procedures.
When making your case, consider weaving the outputs into a compelling story rather than slaying your audience with numbers. Note that we are not suggesting you make up a story. Rather, create a narrative that demonstrates how the results can drive operational improvement or enable innovative new products and services in terms the audience understands.
8. Consider the risks, real and perceived
With machine learning, an individual’s future actions can be predicted with a high degree of confidence, creating the potential for what John Foreman, MailChimp’s Chief Data Scientist, called “laser-guided disingenuous arguments” for targeting and marketing. The ethics, including the privacy vs. value debate, are important and must be addressed upfront.
Another consideration is how our expectation of fallibility changes when a machine decides. As machines increasingly encroach into visible decision-making roles, the human reaction from both your customers and employees must be addressed. Building trust is critical for full adoption and engagement.
9. Proactively adapt business processes
Deploying machine learning systems may require giving up control as you automate certain decisions and actions. Machine learning may also enable development of entirely new products, services or customer engagement models.
Careful consideration must be given to the resulting business implications. Ensure that you can implement insights from machine learning by exploring the effects on existing business process flows and roles early in the process. If you’re automating or augmenting key decision points, clearly identify and explain how these systems will work within the context of human workflows.
10. Plan for ongoing care and feeding
To stay current and deliver results, machine learning algorithms must be continuously refreshed and refined based on data that reflects current circumstances.
If you apply machine learning successfully – and persuade consumers to buy more, use different commercial channels, and turn right not left – the patterns used to create the model will change. Machine learning models need to adjust to account for these shifts in behavior. Your business and development processes do too.
About the Author
Kimberly Nevala is a director on the SAS Best Practices team where she is responsible for industry education, key client strategies, and market analysis in the areas of business intelligence and analytics, data governance, and master data management. Kimberly has more than 15 years experience advising clients worldwide on sustainable business strategies, business-IT alignment and enterprise solution deployment. Her work has been featured by CIO Asia, Knowledge World, Network World and Datanami.
- Article How health care leaders deployed analytics when crisis hitDuring the COVID-19 pandemic, some health care providers were well-positioned to respond to rapid changes in demand. The factor that most distinguished them was that they already had a strong capacity in place for using data to inform decisions. Read about three key takeaways from their experiences.
- Article The untapped potential in unstructured textText is the largest human-generated data source. It grows every day as we post on social media, interact with chatbots and digital assistants, send emails, conduct business online, generate reports and essentially document our daily thoughts and activities using computers and mobile devices.
- Article A guide to machine learning algorithms and their applicationsDo you know the difference between supervised and unsupervised learning? How about the difference between decision trees and forests? Or when to use a support vector algorithm? Get all the answers here.
- Series Meet the data scientist: Kristin CarneyWhen Kristin Carney graduated with a BS in mathematics, she wasn't sure what she wanted to do with her degree. That’s when she began researching data science.