Making machine learning work for marketing: 5 tips
We were honoured to have DAA co-founder and president Jim Sterne join us at the most recent edition of our Digital Analytics Forum event in Paris where he treated our audience to a riveting talk on machine learning in marketing, the topic of his latest book.
During his presentation (which you can watch below), Jim took us through the rise of AI and the benefits of machine learning for marketers, but also highlighted why the human brain and savvy will always be necessary to complement our computers. And because machine learning has certain particularities to consider, Jim also shared tips on successfully onboarding machine learning into your organisation… catch them below!
The 3 types of machine learning
Before taking any action, it’s important to truly understand what machine learning is and how it can be applied. Jim provided a run-down of the 3 distinct types of machine learning and what they can each contribute:
First, there’s supervised machine learning, where the computer receives a massive quantity of labelled data to train on, returns statistics-driven results, and receives feedback and corrections (hence the “supervised” aspect). Jim sums up supervised learning as: “I know the answer and I’m teaching the machine. I’m giving it lots of labelled data and lots of examples. I explain, and it listens.” Take image classification for instance, where the computer has “seen” thousands of images of, say, dogs, and has “learnt” which set of features constitute a dog.
Then there’s unsupervised learning. Like supervised learning, we have vast amounts of data, but on the contrary, we don’t have the answers to train the computer on what’s true or false. So we pour in all our data and ask the system, “Show me a pattern, something interesting, a correlation or an anomaly,” says Jim. And the machine grants our wishes, highlighting curious correlations and relationships between data like “people buy more online when it’s snowing”. But the pitfall to avoid with unsupervised learning is assuming causality where it doesn’t exist. (Does ice cream cause drowning? Watch Jim’s talk below to find out…)
Lastly, there’s reinforcement learning, where there are no absolute answers — just positive outcomes — and we can reinforce certain results, thereby teaching the machine to optimise. “If the outcome is better, give the machine a numeric reward. If the outcome is worse, give the machine a numeric demerit,” explains Jim. He highlights that for reinforcement learning to successfully optimise outcomes, there are 3 things machines require:
– immense volumes of data
– a very specific goal to achieve
– the agency to act or revise their behaviour after fresh data is fed in
Take a moment to consider the infinite potential of an amazingly powerful computer brain which can wrangle a limitless number of dimensions, combinations and possibilities that we as humans cannot even begin to conceive. Who wouldn’t be excited about integrating machine learning into their marketing strategy?
Jim Sterne’s tips on machine learning for marketing
The combination of man and machine talents can lead to great results, but there are a few important things to keep in mind when onboarding machine learning into your marketing actions. Not only will Jim’s advice help you achieve stronger results, but it will also help ensure your machine learning initiatives maintain buy-in and support from your organisation.
Set clear, specific goals
Forget “raise awareness” or “optimise customer retention” — you must be super specific and narrow with your objectives for machine learning to work toward a real result. Proper examples of a specific goal would be “increasing the number of clicks on my ad banner” or “increasing the number of email opens.” Look at what you want to achieve and break it down into multiple concrete actions that machine learning can help you achieve.
Temper management’s expectations
As seen above, even the machine has a learning curve, so your management shouldn’t expect to see amazing results right off the bat. The machine will start with zero knowledge, experience a rather steep and spectacular learning curve, and will then plateau after results have been optimised.
“Be warned”, notes Jim, “First, [the machine] is a dumb as a rock, then it gets brilliant, and then it just becomes average. So what do you do? Give it a new problem, or fresh data. And then it continues to be an amazing, powerful system.” Be sure to set expectations internally so as not to lose support for your machine learning projects if results seem sub-par at a certain point. And always be thinking about how to iterate on your current project to keep your machine learning systems constantly improving.
Data quality is paramount
If you expect accurate and reliable outcomes from your machine learning projects, you must feed your machines with data that’s equally as complete and trustworthy.
“When your data streams hit the data lake, that is where the magic happens,” says Jim. “If this is all clean, consistent and reliable, then machine learning can come in and help you find interesting things in your data lake.”
But for that magic to happen, “you need a data steward watching each data stream,” he notes, to ensure nothing is amiss with tagging and collection. Indeed, while machine learning can do amazing things with quality data, it cannot evaluate or control the quality of the data it’s being fed — human brains are needed to verify that data is clean, accurate, complete, timely, consistent and without bias. (As an AT Internet customer, you’re already in a prime position thanks to the powerful combination of high-quality analytics data and Data Flow to feed your data lake.)
Start with the right tasks
You might have already dreamt up a vast number of project ideas built on machine learning. But certain tasks are better adapted to your computers than others — Jim suggests favouring tasks that are repetitive in nature and low risk, and which involve large amounts of data, such as ranking, sorting, segmenting, clustering, correlating, and anomaly detection — these are all activities where machine learning can beautifully do the heavy lifting.
All that’s left for the analyst to do afterwards is assess the results, determine what looks useful, and continue to explore further. (Take Explorer’s new anomaly detection feature, for example. Powered by machine learning, it alert analysts to suspicious fluctuations which they can then study in more detail.) By allotting the right tasks to machines and humans, both can contribute their talents where they add the most value.
Use your human advantage
Jim’s final piece of advice to us was to embrace what makes us different from computers — and what makes us human: our common sense, reason, emotion, empathy and imagination.
“You need to decide what problem we’re going to solve. You need to decide what data is going to be considered. And you need to decide if the output from the machine makes sense,” he says, adding that thanks to our common sense, we’re easily equipped to run “the smell test” — simply looking at something and knowing that it’s bad, won’t work, or is a poor idea. And despite computers’ brilliance in other aspects, this is something they can never learn.
“What happens between your ears is astonishing, and that’s what you should make use of the most,” he reminds us.
“Computers are incredibly fast and accurate… but stupid. Humans are slow and inaccurate… but we are brilliant.” — Jim Sterne
With all the game-changing developments brought by AI, he says it will be more important than ever to combine human and computer intelligence to build our brands and create systems that can interface with other systems (voice assistants, for example). “Be prepared that marketing is going to change wildly,” he says. “It’s going to get really weird!”
Catch Jim’s entire talk on marketing and machine learning here:
Looking to learn more about data science and digital analytics? Have a look at our other articles:
Originally published at blog.atinternet.com on December 5, 2018.
