Database Reference
In-Depth Information
inhuman KPI. So we're constantly listening to folks to understand how to
properly use data to add value to our service in a way that our users and
ourselves feel comfortable with.
Gutierrez:
Can you elaborate further why you thought it was necessary to
write a book on data science, even though there were already so many great
topics out there?
Foreman:
It goes back to this thing that I'm passionate about that perhaps
a lot of people aren't, which is that it is really easy to get obsessed with
tools—and you have to watch yourself. It is really easy to get the tools to do
all the work for you. But I think it's important that data science practitioners
know what these tools are doing. You don't have to know everything, but you
should have a general idea.
This was really driven into me in grad school. In one class, my professor made
us build an optimization model by hand. We had to do all of the pivots of the
simplex algorithm by hand on paper. It was awful. Anytime you have to do lots
of operations to matrices by hand, it's just a nightmare. He made us do it once,
and after that we never had to do it by hand again. Though it was terrible,
now when I run the simplex method, I know exactly what's going on because
I've done it by hand.
That's very helpful in terms of intuitively understanding what you are doing.
When you formulate a model, you will now intuitively know what you're doing
and what the method is going to do when you call it. If you don't do it once,
if you don't really learn it and internalize it, then I feel like there's always going
to be this secret doubt you're going to harbor. And you're not going to be
able to fully justify what you've done or believe in because there's this magical
incantation you make at some point like, “Call this AI model.” And you don't
really know what it's doing. You might expect too little out of it or you might
expect too much, but you won't really fully comprehend it.
I think a lot of AI models are very dumb. Naive Bayes is the dumbest thing
you've ever seen. So now that I know how it's built, I've actually done one by
hand, I kind of know what to expect out of it because I know what it's doing
internally.
And so the purpose of the topic that I wrote,
Data
Smart
, is that it takes readers
through a bunch of different types of modeling. It takes people through unsu-
pervised artificial intelligence to data mining. It takes people through super-
vised artificial intelligence modeling several times. It takes people through
forecasting, outlier detection, optimization, and simulation as well. And the
way that they go through the modeling is that they do every single model by
hand in a spreadsheet.
