We all know that the data science has been flourishing and changing faces of all imaginable worlds. We’ve seen AI create paintings and music, we’ve seen it drive cars, we’ve seen it diagnose illnesses, and these are just the most obvious, fancy things. Small irrelevant questions, like “What will Susan buy in a supermarket tomorrow?”, seem so non-futuristic and so ‘80s. And indeed, customer behavior did start off with simple product ranking and customer-similarity solutions.
But that was in the ‘80s. Many, many smart people have been working for decades, and in shadow of other flashier AI achievement, they have created models and tools, which can simulate future customer behavior and reactions in disturbingly high resolution.
“Don’t be fooled, maybe it starts with Susan buying a detergent tomorrow but customer behavior is how corporations are built, elections won and migrations triggered. No intention to start conspiracy theory discussion here.”
Don’t be fooled, maybe it starts with Susan buying a detergent tomorrow but customer behavior is how corporations are built, elections won, and migrations triggered. No intention to start conspiracy theory discussion here.
Let’s go over some basics. Classic ML algorithms by their definition try to find patterns in data. Ask a question, define a target variable that describes it, define features (or don’t if you want to go deep) and then tune, tune, tune. That is if you have the data ready and waiting. Usually you will first have to suffer through data collection process. But understanding customer behavior is not really useful on its own. No matter what you are selling, you need to understand your customers in order to make them buy more, otherwise, why bother. We want to understand it in order to change it, and unless you are clairvoyant, you don’t have the data with patterns of changes you are about to cause.
That takes us to the exciting world of experimenting. It all started very reasonably and transparent, people have been doing A/B testing since forever. The digital marketing era widened the playing field. It’s not about which actor you want to see on a billboard any more, it’s about campaign scheduling, and background colors, and language of the text, and the delivery channel… the number of variables just kept growing so people remembered a thingy called Multi-Armed Bandit, which is a very neat solution for handling high number of experiments. And while the number of experiments keeps growing, digital marketing opened another door, called personalization. Obviously when it comes to marketing, one message doesn’t fit all. And in theory, we have the means to send each person a tailored message that will make her or him a loyal customer. The problem is that, we can’t really push creative guys to come up with millions and millions of unique banners. We will have AI doing that soon, but while we wait for AI to become very good at creating banners on demand, the middle ground has to be the segmentation. Luckily, the existing algorithms are very good at clustering, i.e. grouping similar elements and creating groups that differ one from another. Voila, now the designers need only to come up with 5-10 different sales stories, according to descriptions of target audiences that machine gave them, and the machine will take care of the rest – what to show to whom, when and where. And while it always starts with whether Susan will buy the detergent tomorrow, we can now persuade Susan to buy more expensive detergent, or even stop using detergent, or persuade her to go save the dolphins (no offence to dolphins).
Final word, even though we all like to think we are special, and aware of influences and grounded with our principles, the numbers say we are not, at least on average. And in order to increase your sales, the average is all you need.
Comments