How to Move Machine Learning From Possibility To Practice

  • by , Op-Ed Contributor, February 21, 2017
Imagine a world with no spell check -- then suddenly hearing someone say that there were tools available to autocorrect all your spelling in any language on the fly. "Blasphemy!” people might cry, until they tried it.

Marketers today might think the same thing if you tell them that machine learning can improve their email campaigns.

Marketers don't trust machines to do their work. The process of creating custom-rich algorithms and models has a reputation for being too costly and time-consuming, and requiring a team of data scientists.

Analytical modeling has been around forever, but in the past it was challenging to acquire data, and time consuming to build models that certainly weren’t capable of "training" themselves.

Thanks to advances in cloud computing, packaged algorithms and smarter methods of training models, machine learning has evolved immensely. Today it is much easier and faster to predict what customers may (or may not) do next in real time. Machine learning is moving from possibility to practice.



Using Thompson sampling, a scientific sampling method, machine learning helps marketers avoid losing a portion of their customer base on bad variations.

This sampling method combined with machine learning allows you to test traffic in increments as it drips in. But unlike older site testing models, this uses an explore and exploit method versus an A/B or A/B/n testing approach. It will explore in smaller sample sizes and exploit the winning combinations faster. In human words, this means you might have a winning combination that is biased based on the randomization of your segments and send order, or mobile bias in the morning versus tablet bias at night.

Mobile openers might skew a test early in a campaign, and would help an Open Rate KPI you are trying to optimize, but not necessarily a conversion KPI. Imagine a world where you can optimize both, independently, and continue to self-optimize all day long, when the customer is least engaged (branding effect) and most engaged (shopping effect).   

You might think, how do I do this when I need to drop 3 million emails in the next hour?  Well, it doesn’t help you there, as this approach is most applicable to automations, recurring mailings and those that stream continually as an interaction or event happens.

However, you might ask yourself why you’re not using machine learning to help identify customers' daily patterns, versus using the 9 a.m. to 10 a.m. window and reaching only 20% of your customers.  

Take a welcome series or cart abandonment email that is sent 1:1 as an event happens. In the past, testing these emails has been a challenge, given the need for large, statistically significant sample sizes and most companies' inability to run tests side by side on triggers. With this approach, you can test many variables with several different KPIs all in the same stream, such as a series of subject lines, designs and offers, then independently exploit these attributes.

This approach can help determine the impact of day parts as well as interactions or device bias. We know consumer patterns are shifting and it’s OK to optimize open rates on the morning commute, but it may not be realistic to expect a consumer to shop in intelligent ways before coffee.

It’s beginning to sound a lot less blasphemous, right? As author Ray Kurzweil once said: "Our technology, our machines, is part of our humanity. We created them to extend ourselves, and that is what is unique about human beings.” Machine learning is helping marketers form better insights and improve the customer experience. What could be more human?

Next story loading loading..