What’s the hardest part of metric design?
In order to make good data-driven decisions, you need 3 things:
- Decision criteria based on well-designed metrics.
- The ability to collect the data those metrics will be based on.
- Statistics skills to calculate those metrics and interpret the results under uncertainty.
Requirements #2 and #3 have been written about plenty (including by me), but what about requirement #1?
Now that data collection is easier than ever, many leaders feel pressure to drag numbers to every meeting. Unfortunately, in the midst of the feeding frenzy, many of them fail to give metric design the amount of thought it deserves. Among those who are willing to put the effort in, most are making it up as they go along, as if it’s brand new.
Psychology — the scientific study of mind and behavior — has had over a century to stub its toe on the dangers of attempting to measure vague quantities that haven’t been properly defined, so the field has learned some solid gold nuggets that business leaders and data scientists would be wise to borrow when designing metrics.
If you’re not convinced that metric design is hard, grab a pen and paper. I challenge you to write down a definition of happiness that’s so ironclad that no one could take issue with your way of measuring it…
Tricky, right? Now try it with some other abstract nouns people throw around daily, like “memory” and “intelligence” and “love” and “attention” and so on. It’s damned near miraculous that any of us understand ourselves, let alone one another.
And yet, this is exactly the first hurdle that psychology researchers must clear in order to make scientific progress. In order to study mental processes, they must create precise and measurable proxies — metrics — to work with. So, how do psychologists and other social scientists think about metric design?
How do you rigorously, scientifically study concepts that you can’t easily define? Concepts like attention, satisfaction, and creativity? The answer is… you don’t! Instead, you operationalize. For the purposes of this example, let’s suppose you’re interested in measuring user happiness.
What is operationalization?
What is operationalization? I’ve written an intro article to it here for you, but the upshot is that when you operationalize, you first say to yourself, “I am never going to measure happiness and I’ve made my peace with that.” Philosophers have been at this for thousands of years, so it’s not like you’re suddenly going to come up with a single definition that satisfies everyone.
Next, you distill the measurable essence of your concept into a proxy.
Always remember that you are not actually measuring happiness. Or memory. Or attention. Or intelligence. Or any other poetic fuzzword, no matter how grand it sounds to you.
Now that we’re okay with the fact that we’ll never measure happiness and its friends, it’s time to ask ourselves why we even considered that word in the first place. What is it about this concept — in its fuzzy form — that seems relevant and pertinent to the decision we want to make? What concrete (and obtainable!) information would lead us to prefer one course of action over another? (Metric design is much easier when you have actions in mind before you begin. If possible, think about potential decisions before attempting to design a metric.)
Then we distill the core idea that we are after to create a measurable proxy — a metric that captures this core essence we care about.
Define your metric before you name it.
And now comes the fun part! We’re allowed to name our metric anything we like: “blorktibork” or “user happiness” or “X” or whatever.
The reason it doesn’t make sense for us to be arrested by the language police is that no matter how hard we work at designing it, our proxy will *not* be the Platonic form of user happiness.
While it may suit our needs, it’s important to remember that our metric is unlikely to fit everyone else’s needs too. That’s why it would be silly to lock horns in a useless debates about whether our metric does or does not capture True Happiness. It doesn’t. If you’re desperate for some kind of One Metric To Rule Them All, there’s a Disney song for you.
Any metric we create is simply a proxy that suits our own needs (and possibly no one else’s). It’s our personal means to a personal end: making an informed decision or summarizing a concept so we don’t have to write a whole paragraph every time we mention it. We can get along just fine without involving the language police in either one.
So far, so good. You simply determine what information you’d need for your decision, then you figure out a way of summarizing that information in a way that makes sense for your needs (ta-da, that’s your metric), and then name it whatever you like. Right? Right, but…
There is a hardest part to all this. Any guesses as to what it might be? Tomorrow, I’ll share the answer with you — don’t forget to subscribe either here on Medium or on social media (Twitter, LinkedIn) so you don’t miss it. In the meantime, share your thoughts on what the hardest part of metric design is here or here.
If you’re keen to learn more, watch lessons 039–047 from my Making Friends with Machine Learning course. They’re all short videos of a couple of minutes long. Start here and continue in the attached playlist:
If you had fun here and you’re looking for an applied AI course designed to be fun for beginners and experts alike, here’s one I made for your amusement:
P.S. Have you ever tried hitting the clap button here on Medium more than once to see what happens? ❤️
Metric Design for Data Scientists and Business Leaders Republished from Source https://towardsdatascience.com/metric-design-for-data-scientists-and-business-leaders-b8adaf46c00?source=rss—-7f60cf5620c9—4 via https://towardsdatascience.com/feed