Going Deep: Seven Myths About Machine Learning, Continued


Mythbusting continues in Part Two of Stacie Caine Bogdan’s piece on machine learning. A shorter version of this originally ran online at MediaPost

Part Two
(Read Part One here)

Myth Four: Eventually machine learning will determine a one-size-fits-all winner.

This can be a tough one for marketers with backgrounds in disciplines like direct mail – where in simplistic terms, you are working to determine a new winning control.

Think differently regarding machine learning for optimization, where the emphasis is on targeting, personalization and experience. With the machine’s ability to ingest consumer attributes and test multiple experiences, the goal is to determine the best outcomes for each customer type, not a one-size-fits-all experience.

Myth Five: Machines can learn to target immediately.

Building from your hypothesis, think of the first phase of machine learning almost as a random test. By serving different experiences, the machine learns what consumer attributes and factors correlate – and what is effectively engaging customers. Through this experimentation, the machine learns and targeting capabilities improve, advancing optimization.

One way to accelerate targeting: consider limiting or consolidating the number of attributes used in the model. In many cases, some attributes are not relevant: eliminating them expedites performance.

Myth Six: Machine learning takes the place of random A/B split testing.

In the world of machine learning, there is room for both A/B and multi-variant testing. A/B testing may be all that is required when decisions are simple, data is not available in real time or you simply want initial insights before starting more complex testing.

Myth Seven: Machine learning can always outperform.

Machine learning frequently delivers amazing insights and outcomes, but it doesn’t always achieve more. The quality of the inputs – hypothesis, data and experience – are critical to achieve the performance outcomes.

Four common areas where machine performance may go amiss: input attributes are not relevant; too many attributes does not allow machine to reach statistical significance; target audience is too homogeneous; and, creative execution is not relevant to targets.

Back to the initial analogy: while a robotic vacuum is a great tool, it doesn’t replace the need for an upright one. Akin to my last big myth regarding machine learning: that it will eventually replace the need for marketing and analytics experts. Quite the contrary – machine learning simply enables us to be more strategic and empowered as we make human decisions on products, media, positioning and customer experience.

Categories: Connected Data, LFX Conversion Platform, Marketing Insights, Performance Marketing, Performance Media


With more than fifteen years of business intelligence and marketing experience, Stacie Caine Bogdan leads the Strategy and Insights team at Leapfrog, an iProspect company. Combining performance analytics, marketplace expertise and consumer insights, her team regularly improves business outcomes and advances conversion optimization.