A BLOG POST BY NATHAN OLIVER
Programmatic advertising is increasingly being driven by machine learning, and for good reason. Whether it’s good prospecting performance, campaign management or audience optimisation, there are many areas of success. So why would we doubt machine learning?
We just take for granted that good results mean that the model knows what it’s doing.
For context, imagine going to a job interview where the only other candidate is a machine learning model. It’s fierce competition; the model will beat you at almost every task and complete the work in a fraction of the time that you can. But imagine if the statements they were making about their capabilities weren’t actually true, or they were inflating the truth, like people often do in job interviews.
As it turns out, machine learning models are expert liars. They make sure to optimise by any means necessary. If blurring the truth helps them optimise, then they do it without a second thought. Imagine how much an unchecked model could get away with in the fast paced world of programmatic! Artificial Intelligence (AI) researcher, Sandra Wachter, actually calls machine learning algorithms “black boxes” because, “there is a lack of transparency around machine learning decisions and they’re not necessarily justified or legitimate”.
So, how can you ensure that your machine learning model is telling the truth? Well, treat it like a job interview. If you were interviewing a candidate and they made a statement you were sceptical of, how would you respond? Most of us can sniff out a bluff when we see one, it’s just tougher when your bluffer has a steel face. “I have 25% higher performance than any competitor”. “I use all the information I can to sieve the good from the bad”. “I optimise to whatever you want”. Let’s just analyse these statements for a moment, shall we?
I have a 25% higher performance than any competitor.
Great. Everyone wants better performance, and 25% is huge! So how exactly is that measured? Using test data? Simulations? Perhaps an artificial trial? Was there any bias? How much budget was involved? What was the optimal KPI and how did the others measure?
Machine learning technology can be time consuming and expensive, and it’s remarkably easy to waste money on a bad algorithm. Having good, solid proof that a model works is a great way to avoid this and you should always ask for more evidence if you’re unsure on the results. If your model vendor can’t offer you access to an analyst who can back up the numbers with the work, that should be a red flag.
I use all the information I can to sieve the good from the bad.
Love it. Everything at your disposal. Everything treated equally. Every feature as important as each other. Knowing whether someone has bought a product before is just as important as the colour of your socks. Okay, an exaggerated example, but worthy of concern.
If someone tells you they use all the data in the machine learning model, you should feel free to ask how and why. Why is all of the data used? Why is all this important? What tests were run to prove it? Are you even allowed to use all of the data? Everyone’s familiar with the concept and purpose of GDPR, but how familiar are you with how you can use data in machine learning? Always ask the question.
I optimise to whatever you want.
Fantastic! This means I can get exactly the results my client is asking for. I can be as picky as I like and still be successful. I can leave it all down to the model. I don’t need to do anything; the client can choose.
Brands have clear metrics for us to hit and it’s the job of client services together with data engineering to ensure the machine learning optimises towards the KPI. However, the beauty of machine learning is that it frees up the client services team to do more than just achieve the brand’s KPI; they can help brands achieve their business goals too. With thousands of successful campaigns under their belts, client services knows what works and what doesn’t. If you ever hear the phrase “the model just does it”, you should expect to be able to contact a specialist at any time to make sure it’s doing what your clients want.
Lastly, the most dangerous thing to hear when discussing machine learning? Silence.
If you’re talking about purchasing machine learning and your vendor can’t or won’t answer your questions, it’s time to bail. You should feel empowered to ask any and all questions. Just like a job interview, if the answer isn’t a good fit then neither is the candidate.
Not knowing about or not understanding machine learning is perfectly fine. What’s not acceptable, is to say, “machine learning just does it”. Everyone has a right to understand machine learning and we need to be brave enough to open the door.