Numbers are a property of the universe. Once Earthians figured that out, there was no stopping them. They went as far as the Moon.
We use numbers in business and life. We measure, we look for oddities, we plan. We think of ourselves as rational.
I, for one, like to look at the thermometer before deciding if I shall go out in flip-flops or uggs. But I cannot convince my daughter to do the same. She tosses a coin.
More often than we like to think, business decisions are made the way my daughter decides on what to wear. I need an illustration here, so let me pick on workers’ compensation. If you have workers, you want to reward them for good work, and by doing that, encourage the behaviors you want to see more of – you want them to work harder, better, quicker, and be happy. You can measure productivity by amount of, say, shoes sold. You can measure quality by, say, number of customers who came back to buy more shoes. You can measure happiness, by, say . . . okay, let’s not measure happiness. How do you calculate what the worker compensation shall be based on these two measures?
50/50? 75/25? 25/75? Why? Why not? This is where most businesses toss a coin.
Here is an inventory of types of questions people tend to answer by tossing a coin:
- Should you monitor the dollar amount of sales, or the percentage of sale increase?
- Which of the two measures lets you better predict future performance?
- Why would it?
- How accurate are the predictions?
- How big shall the errors be until you feel the measure doesn’t make accurate predictions? Why?
- Which measures shall be combined and looked at together?
- In which way?
- Where would you set up thresholds between good, bad, and ugly?
- Why? Why not?
- If some numbers are way off, how do you know it is an exception and not part of some pattern that you don’t see?
If not tossing a coin, it is common practice to answer these kinds of questions based on a gut feeling. To answer these questions based on evidence instead, there shall be a way to evaluate the gut feeling, together with bunch of other hypotheses, in order to choose a hypothesis that actually true and works. This is hard for humans. Not only because it requires a lot of modeling and computations.
Conceptually, as humans, we tend to look for reasons and explain things. It is hard for us to see a pattern if we don’t see why it works. “I wouldn’t have seen it if I hadn’t believed it” as one wise person said. Admit it, we are biased. We won’t even consider evaluating a hypothesis that looks like a complete nonsense.
Computers, on the other hand, don’t have such a problem. Machine learning can create and test thousands of crazy hypotheses for you and select the best one. That is, the best one in predicting, not explaining. They can also keep updating the hypotheses as conditions change.
That’s why I believe AI is a new BI. It is more thorough and less biased then us humans. Therefore, it is often more rational.
I am fascinated to learn about ML algorithms, and what they can do for us. Applying the little I learned about Decision Trees to the worker’s compensation dilemma above, this is what I get. Let’s pretend the workers get a bonus at the end of the year. The maximum amount of the bonus is based on their salary, but the exact amount is a percent of the maximum based on performance – partially on the amount of sales, partially on the number of returned customers. These are your predictors. Your goal for paying off the bonus is that next year your workers have increased amount of sales AND increased number of returned customers at the same time. That’s your outcome.
Decision Tree algorithm will look at each possible combination of your predictors, and will measure which one better divides your outcomes into categories. (They say it is a division that minimizes the entropy and increases information gain).
Would we try to do that “by hand,” it would’ve taken so much time. But here we have the most effective bonus recipes figured out for us. Some of the recipes may look counter-intuitive; we may find out that the largest bonus is not the best encouragement, or some such. But, again, figuring out “whys” is a different problem.
And here is my little classification of business intelligence tasks that I believe AI can take over and improve upon.
As a human and a designer who welcomes our machine learning overlords, I see their biggest challenge in overcoming our biggest bias, the one of our superior rationality.
Nice posting, Julia!
I agree that we poor humans have a big and undeserved bias toward assuming a superior rationality. But on the other hand, it’s healthy to be skeptical of pronouncements handed down from a black box. Machine learning is only as good as the data that feeds it, and as you and I both know, this is often sorely lacking in both quality and quantity. Garbage in, garbage out.
The challenge, then, for us designers is conveying AI-based recommendations in a way that A) overcomes our emotional biases, and B) satisfies legitimate skeptical inquiry.