HowSocialMediaHelpsTheMusicIndustry
Trung Huynh, Gregory Mead, Matthew Jeery, Jameel Syed
{trung,greg,matt,jameel}@musicmetric.com
Tracking artist popularity, fan interactions & consumption
Musicmetric is a technology start-up based in London, UK that tracks all aspects of music consumption
online including: social network monitoring, web crawling and sentiment analysis, peer to peer download
tracking, in
uential fans identication, fan demographic information and geographic data on where an artist
is popular. Our users include record labels, marketing companies, advertising agencies and broadcasters in
the music industry.
Abstract
Over the past decade, social networks have been
rapidly becoming the most popular platform for
people to communicate and discuss their favourite
artists, songs and concerts. The online footprint
left by music fans is getting larger by the day,
and by tracking this information in real time
we can supply the music industry with valuable
data on artist popularity, consumer behaviour,
fan interactions and opinions. A few examples of
how this information can help the music industry
include: targeted marketing, optimising eciency
of marketing campaigns, discovery of emerging
artists, identifying and targeting \super fans" and
minimising damage from piracy.
In this poster we explore some of the methods used
by Musicmetric to track online behaviour of music
consumers on social networks and the web, and how
this information is used by the music industry.
Distributed data collection
Data Warehouse
Plays
Views
Comments
Downloads
Fans
Mentions
MySpace
Facebook
Twitter
YouTube
Radio
Wikipedia
Last.fm
iLike
Soundcloud
P2P
Web
Figure 1: Data collection process overview
Fan interactions
Statistical analysis is performed on time series data,
allowing patterns and trends to be detected.
Figure 2: MySpace, Last.fm plays over time
The time series data available in the Musicmetric
application covers such data as plays, views and
comments over time.
Regular batched peak detection and cross-correlation
analyses are performed on the back end data ware-
house to pick out patterns in time series data, suc-
cessfully clustering together artists following similar
trends, and allowing regression and predictive trend
analysis on future artist activity.
Analysis that quanties the eects of an event on time
series data, for example the eects of a press release
on social network plays can be calculated using time
series modelling, giving insight into the eects of any
artist related activity.
Fan opinions
Parts of our product are driven by semantic analy-
sis; we do not just show how many people are talking
about an artist, but also their opinions, the sentiment
and common topics surrounding them.
We have implemented a set of statistical machine
learning models that can be trained with dierent
corpora (contexts) so they work well for general lan-
guage but are also much more accurate for the pre
dened contexts - for example, professionally writ-
ten articles, fan comments and tweets are all dier-
ent contexts and therefore have dierent sentiment
analysis models trained for each one. Using this ap-
proach allows our model to become more and more
intelligent as we keep downloading data to retrain it
frequently. The performance of our method is shown
in the confusion matrix below:
Negative Neutral Positive
Negative 96% 2% 2%
Neutral 16% 69% 15%
Positive 5% 6% 88%
Table 1: Sentiment analysis confusion matrix
Fan influence
It is important to rank sources of information, partic-
ularly those derived from text, to accurately show the
real popularity of an artist, and to eliminate spam
or multiple-repeated articles created by automated
spam bots. Musicmetric sources are all ranked for
in
uence using network/graph analysis algorithms
based on the graph eigen centrality combined with
user behaviour proles. This allows sources of buzz
to be scaled by their overall in
uence on the music
related web, thereby giving more importance to an
article published on a high trac website than one
on a barely read blog. A unique aspect of this is the
ability to rank over sub networks related to a par-
ticular genre, and show which websites are the most
in
uential for a particular artist. A high performance
computing architecture is used to reduce turnaround
time for large scale network analysis.
The Musicmetric network analysis algorithms are also
extended to work on social networks like Twitter or
MySpace, eectively identifying the most in
uential
fans of an artist - who would be ideal evangelisers for
any artist related activity.
Fan profiles
Data mining fan information on a large scale allows
us to prole the locations of an artist's fans and their
demographic information.
Figure 3: Fan locations for Lady Gaga
We employ a vector model to minimise errors in
resolving geo-location from un-structured location
string elds.
Figure 4: Lady Gaga's fan age distribution
1