Database Reference
In-Depth Information
Perlich:
My good friend and PhD advisor, Foster Provost, met with me and
told me that he had been working with a startup that was doing targeted dis-
play advertising. He had been their academic advisor for a while and thought
the group was really amazing and super smart. He suggested I have lunch with
them. I went for lunch and the rest is history. So now I am the chief scientist
for what is called Dstillery, though it used to be called Media6Degrees. I have
been doing this for the past four years and have never regretted the move.
Gutierrez:
Tell me about Dstillery's history and focus.
Perlich:
Media6Degrees/Dstillery started out six years ago to focus on dis-
play advertising. Initially, the premise was around social advertising—that is,
being able to target people's friends. The idea was that we could work with
the marketer to observe who is buying their product online. We would then
use the purchaser's browsing history to see what sites they had visited and
compare it to the browsing history of people they were connected to on a
social network. This co-visitation of sites was used to establish connectivity
between people. We would then reach out to those people who were con-
nected to the original seed of people purchasing and show them the mar-
keter's advertising.
We initially started out with MySpace. What we realized pretty quickly was
that while social data was okay as a predictor for certain types of products,
it just did not work for other types of products. Once you got to products
or services that did not have a social connection, looking at people's interests
and the actions they had taken was more useful. Reading a blog on the New
York marathon is actually a much better predictor of the fact that they are
probably a runner, than the fact that they had, at some point, looked at the
MySpace page of a person who was a runner. Being able to find that signal at
scale, that is really what we are doing right now.
Over time, Media6Degrees/Dstillery grew into a very different concept, where
the view was that it actually does not matter whether you are friends with the
person who purchased something. What actually really mattered—if you are
really technical about it—is being able to predict the probability that some-
body is going to buy a product. That is the conceptual view, which is great
because it is what I love to do—solve large-scale machine learning problems.
We do it for financial, travel, and any other type of consumer products/ser-
vices you can buy online across all the world's borders. All that is necessary is
that you have some online touchpoint with your customer. Recently, we have
started focusing more on mobile and videos, so there is a bit of a broadening.
That said, I am not able to advertise milk. So few people buy milk online that it
is basically hopeless. So other than the things not found for sale online, when
we find some kind of a signal that we can track online, then we use it in the
predictive models.
