As the cost of storage and computational power has decreased in recent years, so the speed of development and range of applications for machine learning has increased. As an industry, we’re now talking about how to trust machines to make media delivery optimisation decisions for online advertising. But we constantly stumble when we have to articulate and then implement machine learning versus the tried-and-trusted approaches of targeting.
Consider the design prototypes developed by Arup, an engineering firm, for a component to support cables in an outdoor lighting system. It originally intended this example to showcase the benefits of 3-D printing - another hot business trend over the last few years - but it also shows the use of computational power to optimise component design for superior weight-to-strength performance. There’s an analogy here between the process an industrial designer has to go through to trust the machine and the process the online advertising industry needs to go through.
Bear with me. If you saw an image of the conventional design in isolation, you would have a reasonable chance of guessing what the component was intended for. Once you ask the machine to redesign the component to have the same strength, but a lower weight, while roughly maintaining the position of the cable joins, it looks a little weird compared to the original - but it’s 40% lighter. When asking the machine to redesign the component entirely, it becomes unrecognisable from the original and you have no chance of knowing it’s purpose if you were to see it in isolation. But it’s 75% lighter, while doing the same function. That’s incredible.
As Wired reported, “the biggest challenge had little to do with mastering the machines or exploring the material science of the steel, but rather letting go of old design thinking. Decades of working with traditional materials, using standard design processes and conventional CAD software created habits that were difficult to break.”
In my analogy, we start with traditional, manual targeting selection for online advertising campaigns. We know this works and the insights and delivery reports are comfortingly familiar and recognisable. You can often infer the brand by its targeting preferences.
If we then trust the machine to identify and reach the audiences most relevant to a brand, like the redesigned cable support, it is somewhat familiar. The machine output looks like a list of audiences indexed by relevance to the brand, but the insights can challenge marketing assumptions. For example, a hotel booking service might go from recognisable “hotel intender” and “destination” audiences to finding the trigger behaviours that cause consideration for their service, such as family events, theatre, or shopping trips.
Finally, like the machine-originated design, we learn to remove many of the constraints on campaign delivery optimisation for the machine (though we leave sensible constraints like geo and brand safety in place) to deliver superior results in terms of cost per action and the number of people exposed to generate that result. However, the insights are now disconcertingly unrecognisable to most non-data scientists.
If I play with the words from Wired, the parallels with advertising technology are clear. “The biggest challenge has little to do with mastering the machines or exploring the audience composition of the brand, but rather letting go of old design thinking. Decades of working with traditional segments, using standard campaign processes and conventional campaign software has created habits that are difficult to break.”
If industrial design can move from traditional, welded parts to trusting the machine, advertising can.