i3 - January/February 2018 - 79

Wanted: More Data
Just like your hypothetical best friend,
effective algorithms need to know
a lot about music, movies and TV
shows and a lot about you. In other
words, the algorithms need descriptive
metadata about both content and consumers. Lots and lots of data.
Here, entertainment platforms face
a potential roadblock: how do they
get the deep descriptive information
they need? In practice, they turn to
companies like Gracenote, which
was acquired by data and measurement powerhouse Nielsen in 2017.
Gracenote's metadata powers leading
entertainment and media companies
including Apple, Comcast, Ford Motor
Company and Sony.
Gracenote's General Manager for
Music and Auto, Brian Hamilton,
says the company's data sets are vast
and growing. "We've expanded our
TV genres from just the basics like
comedy, western and crime drama, to
more than 3,000 genres today. On the
music side, we already have very deep
metadata, about 2,000 descriptors
79
2018
C TA .JANUARY/FEBRUARY
tech/i3

around an artist or an album. We've
now taken that to a track level." The
company is also tracking a broad
range of content globally, from overthe-top television catalogues to the
most popular short-form videos on
YouTube.
The amount of Gracenote's user data
is constantly expanding as well. As
Hamilton explains, user data needs to
be more than just what a person consumes and how they browse the web.
"It's also data points about the time
of day and context," he says. "Was the
content consumed on a mobile phone?
What was the time of day? Were you
stressed at the time?"
Hamilton admits that for the
moment, incorporating heart rate
monitor data as a measure of stress is
a futuristic vision. But it illustrates the
scale of Gracenote's ambitions.
SPONSORED

ever recorded. This person would
know both what you like and when
you like it: upbeat pop songs in the
morning, for instance, and slow
country ballads at the end of the day.
As the world's most voracious music
consumer, your friend would also
know every measurable quality of
every song available.
This friend could offer songs specifically suited to your present moment
before you even ask. And if you do ask
for a "romantic" song, for example,
your friend could account for your
favorite genres, your mood, the time
of day, and even the people you're with
before selecting the right track.
Of course, when it comes to content
discovery, these "best friends" are
actually algorithms offering what we
want, when we want it, even when we
didn't realize we wanted it. And if the
algorithms haven't yet achieved "best
friend-level" mastery, they're constantly learning and getting better.

Humans Need Help
Like content discovery, the process of
collecting entertainment metadata
is one that humans cannot effectively manage alone. First, there's the
familiar issue of magnitude. "There are
about 200 million tracks in our music
database today, and the volume of new
descriptors created every day in music
is just astronomical," says Hamilton.
"You literally can't approach music
from a solely editorial perspective."
Another challenge is the number of
possible descriptive data points about
each piece of content.
"Machines are actually better at
tasks that are highly systemized and
complex," says Hamilton. "Listening
to a song for mood is very subjective.
Having hundreds of editors around the
world distinguish moods like 'sensual'
or 'brooding' usually result in very
different song selections. However,
machine learning techniques, when
trained correctly, can be very prescriptive and uniformly analyze
millions of songs across all genres with
a single criteria."
Gracenote's solution for obtaining
property metadata for both music and
video is, therefore, a mixture of man
JANUARY/FEBRUARY 2017

81
http://www.gracenote.com/
i3 - January/February 2018

Table of Contents for the Digital Edition of i3 - January/February 2018
