Leapfrog’s Senior Director of Strategy and Insights dispels common myths on machine learning. A shorter version of this post originally ran online at MediaPost.
Quick analogy: you’re excited about a newly purchased robotic vacuum. To make the most of the new device, probably a good idea to read the instructions and precautions before setting it loose – you don’t want to jam the motor or injure a pet.
The parallel is to machine learning – minus the pet. As someone who works with machine learning and artificial intelligence daily, I advocate its power to advance data-driven, performance marketing. Yet like the new robotic-vacuum, there are guides and limitations.
Let’s start with a definition.
Machine learning is the use of computerized algorithms to analyze large amounts of data; for the machine to learn from this data; and, to make predictions and continually apply learning to new data. All completed by the machine more quickly and efficiently than humanly possible.
In a marketing context, the definition is further refined: to incorporate multi-variant data in real-time; to help teams make the most of data in audience targeting and segmentation; create more relevant customer experiences; and, to make predictions about future outcomes.
One step further: in conversion optimization, machine learning enables personalization at scale. The machine tests and learns what experiences deliver the best performance outcomes – that is, the most clicks, leads, and sales.
The myths of machine learning
As use of machine learning has grown, so have expectations. To make your machine learning efforts more productive, here are seven common “myths” – plus related strategies to get more from your targeting and marketing efforts.
Myth One: You do not need a clear objective.
This one isn’t exclusive to machine learning, but relevant: start with a business objective, a reason for leveraging machine learning. What do you want to achieve or solve?
In the most basic form, the objective is the way to tell the machine what it needs to learn. Having one is critical: reinforcement learning “rewards” the machine for working toward the objective and enables performance optimization.
Myth Two: You do not need to form a hypothesis.
Let’s clear this up immediately: it’s not reasonable to load bundles of customer data into your marketing platform and assume the machine will do the rest.
A more logical and economical starting point is to form a hypothesis. More granular than the business objective, the hypothesis is the proposed explanation and outcome you want to model via machine learning. Or more simply, it is an assumption that you want to test against alternatives.
An example: to optimize user experience, you may want to evaluate landing page features – headline, offer, application, positioning – to name just a few. But imagine the matrix if you were to test every possible variable! Rather, based on subject matter expertise, create a hypothesis to test elements most likely to impact performance based on your data attributes. Of course, the beauty of machine learning is you can test multiple assumptions and alternatives at the same time.
Myth Three: You do not need to calculate sample size and test duration.
This myth is another common “the machine-will-figure-it out” misconception. Like any type of marketing analytics, the sample size must be large enough to have confidence in the statistical significance and the performance results.
How long does that take? The answer depends on the amount of data, number of variables, and the degree of consumer response – and, ultimately, the “learning curve” from the machine itself.
Part two will be published next week…
Topics: Machine Learning