Tip:
Highlight text to annotate it
X
ERIC BRADLOW: Hulu, as most people know, is a
video stream provider.
And we've gotten a data set on 20,000 Hulu users over a four
month period.
And we have information on everything that
they've done on Hulu.
So that includes everything that they've streamed, what it
is they've streamed.
We also know information on the ads that they were served.
We also know whether they pre-roll, mid-roll, or
post-roll ads.
We know how much video they consumed on Hulu for each of
the particular segments or shows.
We know when they stopped watching a show, or in some
sense, clicked off a show.
And so, we also know characteristics of the people.
We also know characteristics of the ads.
So was it a very visual ad?
Was it a more text based ad, et cetera?
And so, the nice thing here is we can now go into the world
of ad optimization.
So we can start thinking about, do you have to do
contextual ads showing--
do you have to match the ad with the show
that someone's watching?
Does it matter the characteristics of the
individual?
So in some way, you have what I call the big three of
contextual advertising, which is characteristics of the
people, characteristics of the ad, and
characteristics of the show.
And how will each of those play out on Hulu?
Now to say that that's kind of our grand vision of what to do
with the Hulu data.
What we've done so far is purely look at what are called
incidence models.
Which means, on given day, did a given person
go to Hulu or not?
And so, all we're trying to do right now is to accurately
match reach and frequency of the Hulu data.
So can we predict these 20,000 people?
Let's say we take the first three months of data.
In this case, it's February, March, and April of last year.
Can we predict what these people are going to do in May?
Can we out of sample forecast the consumption that these
people are going to have?
Now this is hugely important for Hulu, because again--
both towards advertising revenue.
Both towards--
if they're able to find individuals that are likely to
be less interested in Hulu, maybe they can prompt them
with emails.
So in a customer relationship management kind of way, they
can kind of re-prime the pump with these customers.
And so, we've been working very closely with management
of Hulu to try to find out the business problems they're
interested in, and then what we can do.
What we found, actually, is very interesting.
A lot of this type of work has been done historically in
consumer packaged goods.
Let's say someone buying Tide soap detergent over time.
The patterns in interactive media are very, very
different, in the following very specific way.
People tend to go through periods of what I would call
hyperactivity, meaning you're on Hulu, and you're just
watching Hulu over and over again.
And then you go through dormant periods.
Where in traditional models people would say,
Eric Bradlow has died.
Meaning, it's the classic buy to you die model.
Meaning, you go through periods of high activity, and
then all of a sudden, you kind of fall asleep.
The problem with interactive media data
is you wake up again.
And so, you go through periods of hyperactivity, low
activity, and then hyperactivity again.
And that's crucial.
That's very, very different than in traditional marketing
studies on traditional consumer packaged goods.
And so, we're developing a set of models that will allow
people to go hot, cold, hot, cold, cold, cold, hot, hot.
And so those types of models really haven't
been developed before.
And they're very different.
Or if you'd like, we're going to allow for hyperactivity and
clumpiness, which really didn't exist in other types of
models before.