Database Reference
In-Depth Information
Gutierrez: How have you been able to join that point of view with working
at a newspaper?
Wiggins: It's actually completely the same. Here we have things that we're
interested in, such as what sorts of behaviors engender a loyal relationship
with our subscribers and what sorts of behaviors do our subscribers' evi-
dence that tends to indicate they're likely to leave us and are not having a ful-
filling relationship with The NewYork Times . The thing about subscribers online
is that there are really an unbounded number of attributes you can attempt to
compute. And by “compute,” I really mean that in the big data sense. You have
abundant logs of interactions on the web or with products.
Reducing those big data to a small set of features is a very creative and domain-
specific act of computational social science. You have to think through what
it is that we think might be a relevant behavior. What are the behaviors that
count? And then what are the data we have? What are the things that can be
counted? And, of course, it's always worth remembering Einstein's advice that
not everything that can be counted counts, and not everything that counts
can be counted. So you have to think very creatively about what's technically
possible and what's important in terms of the domain to reduce the big data
in the form of logs of events to something as small as a data table, where you
can start thinking of it as a machine learning problem.
There's a column I wish to predict: Who's going to stick around and who's
going to leave us? There are many, many attributes: all of the things that com-
putational social science, my own creativity, and very careful conversations
with experts in the community tell me might be of interest. And then I try to
ask: Can I really predict the thing that I value from the things that the experts
believe to be sacred? And sometimes those attributes could be a hundred
things and sometimes that could be hundreds of thousands of things, like
every possible sequence element you could generate from seven letters in a
four-letter alphabet. Those are the particular things that you could look at.
That is very much the same here as it is in biology. You wish to build models
that are both predictive and interpretable. What I tell my students at Columbia
is that as applied mathematicians, what we do is we use mathematics as a
tool for thinking clearly about the world. We do that through models. The
two attributes of a model that make a model good are that it is predictive
and interpretable, and different styles of modeling strike different balances
between predictive power and interpretability.
A few Decembers ago, I had a coffee with a deep learning expert, and we were
talking about interpretability, and he said, “I am anti-interpretability. I think it's
a distraction. If you're really interested in predictive power, then just focus on
predictive power.” I understand this point of view. However, if you're inter-
ested in helping a biologist, or helping a businessperson, or helping a product
person, or helping a journalist, then they're not going to be so interested in .08
 
Search WWH ::




Custom Search