00:00:01
our guest today is professor Chandra
00:00:03
Hill whose research focuses on social
00:00:06
television which is a combination of
00:00:08
social media and television Chandra
00:00:10
thank you so much for joining us today
00:00:12
thank you for having me
00:00:13
so many people tweet as they were
00:00:16
watching TV shows why is this so
00:00:19
important to the media industry both is
00:00:22
part of figuring out how many people are
00:00:25
actually watching the show and also why
00:00:28
does this matter to advertisers so there
00:00:31
are a number of reasons why it's
00:00:32
interesting and also important the first
00:00:34
one is that for TV shows they can
00:00:37
observe in real time immediate response
00:00:40
and engagement to the content in the
00:00:42
show whether that be sort of organic
00:00:45
response to just the script or TV shows
00:00:49
actually incorporating social media
00:00:51
content and the shows and asking people
00:00:52
to do for example voting while watching
00:00:55
TV but then in addition to looking at
00:00:58
real-time engagement one could also look
00:01:01
at the viewers and estimate things like
00:01:04
demographics interest and get a sense
00:01:07
for who the people are that are actually
00:01:09
watching the TV show so the Nielsen has
00:01:11
just announced that it will start
00:01:12
providing demographic data why is this
00:01:15
notable and or is it really notable well
00:01:18
we've been estimating demographics from
00:01:21
tweets for for quite a long time as the
00:01:25
tweets relate to television viewers and
00:01:27
television content so while it's notable
00:01:30
in that it's an added service for for
00:01:33
Nielsen customers it's something that
00:01:35
can actually be done with publicly
00:01:37
available data for free and what's nice
00:01:41
about it is that if you have a
00:01:44
methodology to infer demographics of
00:01:46
groups or individuals you can do so for
00:01:48
a really large number of people and
00:01:50
Twitter handles so Twitter handles
00:01:52
represent shows and brands and so it's
00:01:55
nice about the ability to actually infer
00:01:57
demographics is that you can do so for a
00:01:59
wide range of Twitter handles more so
00:02:02
than the popular televisions that
00:02:05
television shows that Nielsen typically
00:02:07
follows can you tell us a little bit
00:02:09
more about your approach and how it
00:02:11
differs from what Nielsen has
00:02:13
announced that it's going to do well so
00:02:15
our approach works in the following way
00:02:17
so we start with Twitter handles so
00:02:21
usually we focus on television shows and
00:02:23
brands but the Twitter handles can
00:02:26
represent anyone for example I could use
00:02:28
the Twitter handle of my account and we
00:02:31
start with those twitter handles and for
00:02:33
the Twitter handles we grab the
00:02:34
followers of the Twitter handle for
00:02:37
those followers we grab their tweets and
00:02:39
so each Twitter handle then in our
00:02:42
method is represented by all of the
00:02:45
tweets of all of the followers so you
00:02:47
can think of this as for a particular
00:02:48
show we have all of the follower tweets
00:02:51
not just tweets about the show but
00:02:53
tweets about their daily lives once we
00:02:56
have this document if you will of all of
00:02:59
the tweets of the followers of a
00:03:00
particular handle then we basically
00:03:03
create what's called a bag of words so
00:03:05
think of it as just creating you know
00:03:07
one big vector of words and the
00:03:10
associated counts with them we then
00:03:13
correlate the count so we've normalized
00:03:16
the counts in a special way but we
00:03:18
correlate the counts on these words with
00:03:21
aggregate level demographics of the
00:03:24
shows so we get this data in an
00:03:27
interesting way from Facebook through
00:03:30
the Facebook advertising API which
00:03:34
allows us to get estimates of the
00:03:37
aggregate level proportions of people
00:03:41
that follow a particular thing it could
00:03:43
be a television show like I mentioned or
00:03:45
a brand or a person and what we do is
00:03:51
correlate the proportions of people that
00:03:54
follow a brand or television or person
00:03:56
with these words and find the words that
00:03:59
are correlated with different
00:04:02
proportions of demographics so examples
00:04:04
of those demographics that we've looked
00:04:06
at our gender age education level but
00:04:11
then in addition to that Facebook also
00:04:13
allows you to target different interests
00:04:17
so people with different interests so we
00:04:19
could even look at estimates of the
00:04:22
population for people that like
00:04:24
gardening for example or cooking and so
00:04:26
it's fascinating
00:04:27
even at this aggregate level where we
00:04:29
take the words associated with these
00:04:31
shows the people that you know the words
00:04:33
associated with the people that follow
00:04:34
the shows and these aggregate level
00:04:36
demographics we can do a really good job
00:04:39
when we build models to correlate the
00:04:41
words in the demographics at predicting
00:04:43
the demographics of held out groups of
00:04:47
people and so why this is so powerful is
00:04:50
that you don't have to restrict yourself
00:04:52
to just popular things that you know
00:04:55
perhaps a company like Nielsen would
00:04:56
typically have estimates for a reliable
00:04:59
estimate you can make estimates for just
00:05:02
about anything that has a group of
00:05:04
people that talk about their daily lives
00:05:07
that you can sort of isolate this group
00:05:09
so speaking of that you have you have
00:05:11
something that you call the taco graphic
00:05:14
profile explain exactly what that is so
00:05:17
we've just coined this term to basically
00:05:20
mean that people are what they say so
00:05:24
groups and individuals you know sort of
00:05:27
used specific language on Twitter and on
00:05:30
social media and basically they are what
00:05:32
they say so words that people use are
00:05:35
highly indicative of both their
00:05:37
demographics and interests the needs of
00:05:41
service offers demographic data to
00:05:43
industry but your approach takes that
00:05:46
one step further and actually makes
00:05:48
recommendations based on the data how do
00:05:51
you do that and why is it different than
00:05:54
traditional recommendation systems right
00:05:56
so we have built a recommendation engine
00:05:59
on top of what we call these profiles
00:06:02
for shows and so we wanted to show that
00:06:04
these profiles actually had value and so
00:06:08
what we do is simply calculate the
00:06:10
similarity between shows or really
00:06:12
Twitter handles based on the words that
00:06:15
people who follow the shows say so in
00:06:18
doing so we can calculate the similarity
00:06:21
between anything any Twitter handle and
00:06:23
when we have a new set of users we can
00:06:29
then sort of give one of the shows or
00:06:31
Twitter handles that that user follows
00:06:33
into our big correlation matrix of items
00:06:36
or shows and ask based on the similarity
00:06:40
between the
00:06:41
item that we give our system in all of
00:06:43
the calculations that we've done what
00:06:44
are the things that that user would most
00:06:46
likely follow and we find that
00:06:49
calculating the similarity between shows
00:06:51
in this way just using the words is is
00:06:54
highly predictive when we sort of build
00:06:57
this recommendation engine predicting
00:07:00
what people will follow and so the nice
00:07:01
thing about this is while there are a
00:07:04
lot of strategies for building
00:07:05
recommendation engine so for example
00:07:08
using the product network or the network
00:07:12
of Twitter handles that form by looking
00:07:14
at Twitter handles that are commonly
00:07:16
followed by a lot of people and using
00:07:20
the text it means that you don't need
00:07:22
those networks so while Twitter has a
00:07:25
large network of users there are a lot
00:07:27
of websites that don't write and so here
00:07:30
it's saying that perhaps the the tweets
00:07:33
or text can be used as a substitute for
00:07:36
those network data when it's not there
00:07:38
and when it is there it can be used to
00:07:40
complement it in addition to kind of
00:07:44
showing the value there it also helps
00:07:46
with what's called a cold start problem
00:07:48
for recommendation engines and basically
00:07:51
yeah so this problem is when you have an
00:07:54
item or a product or in the case of TV
00:07:57
as we're talking about now that doesn't
00:07:59
really have that many followers then you
00:08:05
you're not going to recommend it right
00:08:06
so it's quicker to get these tweets you
00:08:10
only need a few followers to start you
00:08:12
know calculating the similarity between
00:08:14
the twitter handle's
00:08:15
so it'll be more likely to be
00:08:17
recommended sooner and it also tends
00:08:21
towards making more diverse
00:08:22
recommendations so as opposed to making
00:08:25
sort of more popular recommendations in
00:08:28
terms of our methodology that we've got
00:08:30
so can you give me an example of a
00:08:32
specific instance where you've been able
00:08:34
to make some recommendations using a
00:08:37
recommendation engine and others also so
00:08:40
we test this on Twitter users and so
00:08:44
we're assuming that the things that
00:08:46
Twitter users follow a representative of
00:08:48
the things that they're interested in
00:08:50
so we basically build our models on one
00:08:52
subset of users
00:08:53
and then make predictions on a holdout
00:08:55
set of users so we've done this in the
00:08:57
context of television shows reliably and
00:08:59
then we've also done this in the context
00:09:02
of brands so so we've focused mostly on
00:09:06
television shows and brands but because
00:09:08
we're using all of the words that people
00:09:12
say without going in with any ontology
00:09:15
meaning you know some list of words
00:09:18
which would be a lexicon or
00:09:20
like relationships between known words
00:09:22
that are meaningful for television or
00:09:25
brands our approach is generalizable so
00:09:27
we take all of the words that people say
00:09:29
without restricting them in any way and
00:09:32
what that means is we can calculate the
00:09:34
similarity between any two things not
00:09:36
just television shows not just brands
00:09:38
and so and it makes their approach very
00:09:42
flexible that's a good company or a
00:09:44
brand use your approach in-house and if
00:09:46
so how would they go about it absolutely
00:09:48
I mean so what's nice about the approach
00:09:51
that that we've developed is it relies
00:09:54
strictly on publicly available data so
00:09:57
which means the data are free so we've
00:10:01
tested our approach on Twitter users
00:10:04
that have revealed their preferences by
00:10:06
following there would still be this
00:10:08
extra step needed for a firm to test it
00:10:11
in their particular context but if they
00:10:14
are trying to just infer the
00:10:16
demographics of Twitter users then it
00:10:18
would work that way but if they wanted
00:10:19
to use the approach for making
00:10:21
recommendations in their context they
00:10:23
would have to test it against their own
00:10:25
users now TechCrunch story about
00:10:28
Nielsen's new offering says the company
00:10:31
has found it while there are a
00:10:32
significant number of people tweeting
00:10:35
about TV and even larger number are
00:10:38
consuming that content and sometimes
00:10:40
those consumers shed new light on what
00:10:43
type of demographic groups watch a
00:10:45
particular show have you seen that in
00:10:47
your research and why is it so important
00:10:50
to tease out passive tweet readers along
00:10:53
with people who are actively creating
00:10:55
content right so we haven't we haven't
00:10:57
focused on that distinction mostly
00:11:01
because we decide that somebody's
00:11:05
interested in a show based on the fact
00:11:07
that they fall
00:11:07
that show not based on the fact that
00:11:10
they're tweeting about it and so all of
00:11:12
the users that followers show would be
00:11:14
in our data set most those that actively
00:11:16
tweet and those two that don't we could
00:11:19
easily compare them my guess is that
00:11:21
they're not extremely different but
00:11:23
perhaps they are but we could compare
00:11:25
them pretty easily so the nice thing
00:11:27
about that would be like if there are in
00:11:29
fact differences then it would provide
00:11:31
insights to a company but we focused
00:11:35
mostly on you know of the people that
00:11:37
follow your brand can we infer or
00:11:39
tv-show can we infer the overall
00:11:41
demographics but it would be easy
00:11:42
because our approach infers information
00:11:47
for groups of users to infer the
00:11:50
demographics of the subset of people
00:11:53
that tweet and the subset of people that
00:11:55
don't for a particular show finally one
00:11:59
last question people both in and outside
00:12:01
the media industry have complained for
00:12:04
years that the traditional system of
00:12:06
Nielsen ratings doesn't accurately count
00:12:09
how many people are actually watching
00:12:10
the show what are the stakes here and
00:12:13
why is doing that so important how can
00:12:16
social media be a game changer in this
00:12:18
so the main thing is that the way that
00:12:22
you can easily collect the data and sort
00:12:25
of watch viewers enables us to watch a
00:12:29
larger number of viewers for free and
00:12:32
therefore make inferences about a larger
00:12:35
number of users and TV shows for free so
00:12:38
most of the criticism of Nielsen data
00:12:40
there there are many but meant one of
00:12:42
the main ones is that there's not a lot
00:12:44
of coverage for niche shows or shows
00:12:47
that aren't that popular because the way
00:12:49
that historically they've collected data
00:12:51
is based on a relatively small panel of
00:12:54
users that have a device in their home
00:12:57
to track their viewing patterns and so
00:12:59
what social media does is it opens up
00:13:02
the space of users to pretty much
00:13:04
anybody that's that's tweeting and so
00:13:07
you're not restricted then to infer only
00:13:10
the demographics of popular shows
00:13:12
because there will be coverage for all
00:13:14
shows and all things on Twitter now say
00:13:17
that with the main caveat that not
00:13:19
everybody's on Twitter right
00:13:21
so there's going to be a bias of course
00:13:25
if you use sort of only the social media
00:13:28
data so while you can make inferences
00:13:30
for a larger number of shows you're
00:13:33
going to be biased and you'll have to
00:13:34
correct for that bias of the fact that
00:13:36
not everybody is on Twitter it's skews
00:13:39
younger and that has to be accounted for
00:13:41
but the promise is that you know you can
00:13:44
make inferences cheaper faster for more
00:13:47
people and for more shows would you like
00:13:50
to say anything about a question that I
00:13:51
should have asked but I haven't so I
00:13:53
think maybe just talking about sort of
00:13:57
the future I'm producting maybe what all
00:13:59
this means for the TV industry so it's
00:14:02
nice to see Nielsen and other companies
00:14:06
beginning to combine different types of
00:14:08
data to make more comprehensive pictures
00:14:12
of their their viewing audience but what
00:14:16
would be nice to see from companies like
00:14:18
Nielsen our partnerships with data
00:14:23
having partnerships with people that
00:14:25
have different types of data that the
00:14:27
masses couldn't otherwise get so with
00:14:30
Twitter you know sort of inferring
00:14:31
demographics is something that we do
00:14:33
it's something that a lot of researchers
00:14:36
now are starting to do for various
00:14:38
reasons and so that's kind of easy what
00:14:42
would be nice to see would be now you
00:14:44
know perhaps customers wouldn't want
00:14:46
this but to see them do things like
00:14:47
partner with credit card companies and
00:14:49
partner with companies for which it's
00:14:52
really difficult to get that data and
00:14:54
see sort of what insights can be drawn
00:14:56
about TV viewers from combining data
00:15:00
from disparate sources that you know
00:15:02
aren't easy to cut genres thanks so much
00:15:06
for speaking with us today
00:15:24
you
