Machine Learning: A Marketing Strategist’s Perspective >

 

Buzzword warning: Artificial Intelligence and Machine Learning. Your ears pricked up a little, didn’t they?

And no wonder. What makes machine learning so appealing is the idea that a computer can help make the optimization decisions that a small or overwhelmed marketing team cannot. Better results and increased ROI with reasonably less effort? Yes, please. Every digital marketer knows they want it (need it!) but that doesn’t mean they know how to use it, or even what it is.

So let’s talk, in broad terms, about how machine learning works and go over a few examples of practical application for large enterprise email programs.

 

What Is Machine Learning?

Machine Learning is a subcategory of Artificial Intelligence in which we (humans) give computers (machines) the ability and autonomy to learn and make decisions based on data. To keep things simple(ish), we’ll talk through three types of machine learning: supervised, unsupervised and reinforcement learning algorithms.

 

Type 1: Supervised

Supervised machine learning algorithms require feeding the machine defined or labeled variables which the computer rationalizes to make predictions. Supervised learning has two primary functions – regression and classification.

Regression relies on numbers. For example, you may feed the machine a subscriber’s total number of purchases with your business, total spend, and the quantity of items purchased. These data points are grouped across a large population of subscribers and the machine can predict the lifetime value of a subscriber. This information then allows you to tailor your messaging to subscribers based on propensity to spend.

For example, if you’re trying to anticipate the ROI for an email campaign, and you have past results to use for feeding the algorithm, you could put in the total number of items sold, the total number of customers who converted and their average spend to predict the net amount for your new campaign.

Supervised learning algorithms can also help with classification. If you have clean, complete data, a supervised learning algorithm can learn what {something} is and find like {somethings} that need to be tagged or targeted. Clean is the operative word here. If you introduce “dirty data,” then the supervised learning algorithm can get confused and fail to produce valid results.

 

Type 2: Unsupervised

Unsupervised learning algorithms creates patterns and clusters from unlabeled or undefined data. When you have a lot of untagged data, machine learning allows the the computer to try and cluster like fields. Unsupervised learning, for email marketers especially, is helpful in the creation of segments that you may not have known existed. Unsupervised learning enables the identification of segments based on a single variable or a set of variables.

For example, the machine may notice that engaged subscribers are those who open while unengaged subscribers are those who don’t. Easy enough. However, it may also notice that “hero” customers open and click once per week, visit at least 6 pages on the site, and make at least one purchase per month while “lapsing” customers do not. Now you have the makings of a sophisticated 1:1 marketing program.

We know that providing more personalized content creates greater results. Often, however, we don’t have profiles on hand that highlight what our subscribers may be interested in. If we take data from all touchpoints (website, email, paid media, etc.) and see how our subscribers interact, an unsupervised learning algorithm can identify segments for us and pave the way to a more sophisticated 1:1 omnichannel marketing program.

 

Type 3: Reinforcement

Reinforcement learning algorithms give the machine autonomy to make strong rules for what it considers success and then incentivize the machine to achieve it.

Within your email newsletter campaign, you probably serve the most recently published content first. But it may be a wiser strategy to use machine learning to identify the types of blog content that resonates best with individual audience segments. The machine is allowed to select articles and success would be defined as engagement. If the audience engages, the machine views the success as a reward and keeps delivering that type of content. If the audience does not like the content, the machine learns that too and will serve other items. Sometimes the machine gets to pick at random to see how the audience engages.

Reinforcement learning is also great for subject line testing. While the good old 80/20 split tests still exist, they’re static. The next subject line will still just be competing against one other. Reinforcement learning algorithms can test one subject line, reward itself for increased open rates and continue to improve. And good companies can write those subject lines in your brand’s voice. Machine learning outputs don’t have to be robotic. Moreover, they shouldn’t be. Good machine learning algorithms will get better and more like your own brand voice.

 

How to Get Started

Step 1: Identify the questions you want to answer. 

A good AI professional will turn anyone away who simply says they “need AI” because they want it. Without knowing the questions you want answered, you may receive results that are unusable. And, it’s hard to go at this alone. While it’s true marketing is a mix of art and science, this truly is science.

We at BrightWave think this is great. Turning data into deliverable results is our shared passion and in years past, we only had the human mind to help us get there. With AI, there’s the real possibility of having help answering questions that were previously impossible to answer. And sure, the fact that machines are getting smarter and smarter can be scary. But it’s also exciting.

Once you identify what questions you want to answer, create a hypothesis showing ROI in order to get executive buy-in and develop a plan for implementation.

Step 2: Determine the kind of algorithms that will help answer your questions.

Now that you’re clear on what you want to answer, identify what kind of algorithms will allow you to answer your questions. Lean on your analytics department— they’re a great resource for helping you determine the appropriate algorithms!

Step 3: Go on a data scavenger hunt.

The last step is meeting with your IT department and other marketing departments. Once those teams are informed, regroup with your analytics team.

 

Last piece of advice: don’t be scared. Start with some low-risk opportunities to prove results and then build out a more robust program. Even if you can’t hire 15 advanced analytic graduates in the next 6 months, you can run one test. One test run, maybe with small results, but they’ll be incremental results that keep improving. Or, one test run and you find results you can’t do anything with, yet, but that’s just a “yet.” There will be more opportunities to use whatever knowledge you gain. And, that test will help you get better at asking the right questions. Fail fast has always been something of a mantra (cliche?) for marketers, but who among us have really tried? Who’s in? I am.

Leave a Reply
(will not be published)