Machine learning may sound expensive and out of reach, but it needn’t be. Nearly every major machine learning implementation has been made available. Open source platforms include offerings from Amazon, Google, Microsofit, Baidu, and many more.
These “kits” substantially reduce the knowledge base required to apply machine learning, to the point it can be nearly turnkey. In fact, machine learning has become so accessible that even the most technically unsavvy of reporters are trying their own experiments.
Thus the barrier is no longer obtaining the PhD scientists required to build a bespoke machine learning application—although some big, moneyed players like Uber are doing just that—it’s about using the tools on the market in ways that truly drive value for the business.
- Ask a great question. In any business, there are important questions that haven’t been answerable with existing methods. For example, which “types” of customers require the least maintenance? If you could identify these groups and bring on more individuals fitting the criteria, it could help maximize ROI. This is a great question for machine learning.
- Look at historic data. For example, which lead generation techniques are associated with more refunds? Getting started can be as easy as finding some historic data relevant to the question.
- Start small. Many enterprises think they need huge data lakes for machine learning, but that’s not true. It takes less data—and less time—than many people expect to use a supervised machine learning system to generate useful predictions.
- Use clean data. If each marketing campaign is tracked differently, the data is “dirty.” It’s better to use a small set of clean data than big data that’s a total mess.
- Avoid decayed data. Another reason not to “go big or go home” with machine learning—data from this month is much better than data from last year. Although there can be reasons to look at longer timeframes, such as exploring seasonal variations, the newest data fitting the problem is the most useful.
- Simplify, sometimes. More attributes are great, but they don’t always lead to better, more applicable results. They can invite greater variation and more error as well. It’s important to know when to use fewer data points.
- Fulfill the need for speed. The less time data sits before it is fed into machine learning systems, the better. If a transaction happens and can be automatically leveraged, systems can return real-time predictions based on current conditions.
- Create a service. Rather than each team hiring its own data scientists, give the entire enterprise access to a service, so teams can access the toolbox for a variety of needs. This could look like a web interface with a curated algorithms and APIs.
- Focus on the concepts. Getting internal buy-in will require that nontechnical stakeholders “get” machine learning. That means understanding what it can do and what problems it can solve, not the math behind the algorithms. Google is an example of a business bringing all types of personnel in on machine learning, at different levels of sophistication.
- Don’t apply results blindly. Machine learning methods are often described as more “Zen” than traditional programming. This inscrutability, however, does not mean what comes out of an algorithm can be put in action without further consideration. Questioning the data inputs, evaluating results against expectations, and determining the proper response are equally important parts of the process.
- Don’t leave it to the nerds. Creating and tuning a model requires involvement of the business experts, who will have a better sense for potential bias in the data and the implications of the insights being sought. This can involve a “translation,” from numbers to narrative to interact with various audiences in terms they can relate to.
- Access expertise. Statistical reasoning and scientific analysis capabilities are vital to avoid turning machine learning into “magic” and applying it in an uncritical manner. If the capabilities don’t exist in house, bring on more experienced professionals, whether permanently or on a consulting basis, to help guide initial projects or big efforts.
- Establish policy. Machine learning can raise new ethical concerns. Just because one can dig deeply into data, should the enterprise do it if it unearths information with privacy implications? Should the enterprise act on recommendations customers might find “creepy” if they heard about? Additional decision-making may need to be applied to avoid missteps.
- Continue the learning. Machine learning is not a “one and done” process. If insights are applied and have an impact, the system has changed and the models may need to be adjusted. Expect to make ongoing investments of time and talent to continue to reap value.
In short, machine learning is more accessible than many believe, requiring less data than many IT pros might expect. And it’s not a whole lot different than other technological tools. To work for the business, it needs to be built into the existing processes, culture, and oversight/governance structures within the enterprise.