BIG DATA = BIG

DISRUPTION?
E NUOVE FRONTIERE DEL
BRAND ENTERTAINMENT

Catarina Sismeiro
Imperial College London
Associate Professor of
Marketing

OVERVIEW

1. Big data
2. Branded Entertainment and how it
can benefit from big data analytics
3. Applications and examples

BIG DATA?
DATA THAT REQUIRES
NEW SKILLS AND A
EW WAY OF LOOKING
AT STORAGE AND
PROCESSING

Big data describes datasets that are so larg

complex, or rapidly changing that they push the ve
limits of our analytical capability. It's a subjective ter
What seems big today may seem modest in a fe
years when our analytic capacity has improved
Joel Gurin, author of Open Data Now

What are the typical challenges?

High Volume
High Velocity
Extremely Unstructured

Not Just associated with internet

(Old Data can be very big!)

New data available in a world of allthings-digital

(e.g., text, image, voice/sound, and video

GOOGLE FLU
PROJECT
A BOLD IDEA TO
SHOWCASE THE
POWER OF BIG DATA
ANALYTICS

Black Box the system does not try to understa

why but wants only to predict
When something changes in the market, predictio
can go wary.
More than just Big Data need Intelligent Data

GOOGLE FLU
PROJECT
A BOLD IDEA TO
SHOWCASE THE
POWER OF BIG DATA
ANALYTICS

Black Box the system does not try to understand

why but wants only to predict
When something changes in the market, predictio
can go wary.
More than just Big Data need Intelligent Data
NOW CLOSED

Thank you for stopping by.

Google Flu Trends and G
Dengue Trends are no l
publishing current estimates
and Dengue fever based on
patterns. [] It is still early da
nowcasting and similar too
understanding the spread of d
like flu and dengue we're e
to see what comes next. []

Sincerely,
The Google Flu and Dengue
Team.
https://www.google.org/flutr
about/

BRAND
ENTERTAINMENT
GREAT
OPPORTUNITIES AND
GREAT CHALLENGES
IN TODAYS
NETWORKED SOCIETY

Not new but new opportunities for content creati

and distribution
Power of amplification and serendipity of soc
networks (emerged more as a platform of us
distributed content than one of user generated conte
Great potential and but also significant risks

HOW CAN BIG DATA

ANALYTICS HELP
BRAND
ENTERTAINMENT?

Start simple and reuse existing tools and analyses (8

of the gains from the initial 20% of effort)
Think of the relevant unit of analysis
(e.g., if I make decisions weekly should I have an ho
demand analysis?)
Combine methods

Discovery and Sensing

Measurement and Prediction
Targeting and Personalization

APPLICATION
PREDICTING VIRALITY AND
HELPING DESIGN
USING TEXT ANALYSIS

TEXT ANALYSIS:
SIMPLE EXAMPLE

Calzedonia Ocean Girls Sono

Stupende!!!
After stemming and after removing
extremely common and extremely
uncommon words, retained only:
calzedonia, ocean, girls,
Stupende
Encode as TextVector (word_id : count)
(5:1,7:1,142:1,846:5)
Create word dummies to use in models

Message

Word 1

Word 2

Word

0 Word not present

1 Word present

Stemming example: eater, eating eat

This post would include one mention to o

monitored brand. The same can be done w
longer texts (e.g., article) like in the vira
study

GOOGLE
T R E N D S D ATA

FITC
H

FITCH
THE
HOMELE
SS

ABERCRO
MBIE AND
FITCH

CAUTION!
BEYOND VOLUM

SENTIMENT AN
MEANING

T he apparent

increase in bran
interest was no
necessarily positiv
Sentiment Analys
and Content Analy
(i.e., what is said a
if it is positive o
negative) is very
important

But remember th
80/20 rule

FINDINGS
CONTENT VIRALIT
DEPENDS ON:
Position of the content
(top of the list)
Practical utility of post
EMOTIONS

Content associated with h

arousal positive (awe, surp
and humour) or negativ
(anger or anxiety) emotio
more viral

Content that evokes low

arousal, or deactivating
emotions (sadness) is less v

EXT MINING AND SENTIMENT

ANALYSIS
MEASURING BRAND
STRENGTH (IFEGY ATLAS)

EXTREMELY POWER
AND VERSATILE

Free or commercia
software perform te
mining for differen
languages

Data can come fro

blogs, online
conversations in so
media, news, ads
emails, commercia
offers, reports

Think of the potenti

What are people
discussing after
watching the last
episode of RDS
Academy?
Discover ideas for t
next episode? Disco
what worked and w
did not?
What content on yo
social media chann

APPLICATION
TARGETING MOBILE USERS WITH
MULTIMEDIA MESSAGES: POWER
OF TEXT AND IMAGE

blem

Mobile phone company wishes to target

ustomers with alternative messages
Messages include offers from many
ategories and combine TEXT + IMAGE+
RICE
Difficulty in learning using traditional testing
methods due to the large scale of offers and
he reduced time to learn (e.g., most expire
quickly not enough testing opportunities)

Solution

Statistical methods (SVR), text mining (simp

word count), and automatic processing of
images (Textons) to predict performance f
different users
Given limited testing possibilities, test only
most promising offers as per the model
Practical issues of deployment that lead to
adaptive system (control vs. testing groups

15

MAGES CAN BE
ECOMPOSED INTO
OLOUR COMPONENTS

e digital representation of images:

Pixels (a structure that provides a
location; could be similar to a Cartesian
grid though alternatives exist)
Colour encoding (e.g., RGB)

om their digital representation we can

process images using a variety of
ers, extract features (e.g, extract text
extract components, determine how
uch vegetation, or how much skin)
nd, considering their overall look at
el, classify and group similar images

IMAGE
NALYSIS
:
PIXELS
ered
values are then
AND
ed (e.g., K-means) and
ILTERS
entroid
of each cluster

Pixel

becomes a visual
word (Texton)

e then describe each

ge with respect of how
of each visual word it
contains

also possible to extract

fic image components

Each image is made of pixels

(similar to a coordinate, a location)
associated to a specific value (e.g.,
in RGB colour encoding)
For simplicity we are considering
here a rectangular grid (this is not
the only possibility)

Examples of Filtered Images

There are many different filters a

some allow the classification on t
overall look and feel of the imag
others try to extract components
the image or identify edges an
ridges

AUTOMATIC
PROCESSING OF
VISUAL
CONTENT
Automatic image
clustering works very
well in finding similar
types of images
The type of image
being used can then
be entered as a
variable in models of
performance
Or develop system to
predict potential
outcome

There are
many spec
tools and m
different wa
of
implementi
image mini

Some are
supervise
and requi
some hum
input and
others ar
complete
unsupervis

The 80/20 r
applies als
here
(start simp

RESULTS

1.5

Price

1.4

Revenue Lift

rmation on images and

xt help to significantly
mprove performance
me methodology can be
plied to devise a system
test different messages
formance metrics could
ude clicks, views, shares

Price and Te

1.3

Price, Texton
Text

1.2
1.1
1.0
10

15

20

25

30

35

40

System Testing Constraints

45

50

55

APPLICATION
REAL TIME MARKETING: SENSING
AND MONITORING

MARKETING

THE STRATEGY AN
PRACTICE OF
REACTING WITH
IMMEDIACY IN DIGIT
CHANNELS TO
EXTERNAL EVENTS
AND TRIGGERS

OUTCOMES *

81% increased custome

engagement
73% improved custome
experiences
59% increased conversi
rates

Challenges: identifying
opportunities that are
adequate for your target
testing quickly the creati

HUMOUR
OPPORTUNITY
REACTION

Sensing can be done b

humans or by automat
systems

Sensing systems can r

on conversation data
among the target
consumers of your prod
(specialized) and set
automatic warnings

Creative can be teste

using automatic system
similar to the one
developed for the previo
example relying on ima
and text

Sensing also useful for

content website (e.g., ne

APPLICATION
WHAT IS THE VALUE OF A SOCIAL
MEDIA CONVERSATION FOR A
BRAND?

blem

esearch has shown that companies can

ffectively stimulate online conversations
hrough innovative campaigns, and that these
onversations increase product sales
How much to invest in these activities depends
n much conversations are worth in terms of
evenue they generate
How much is a conversation in different social
media worth for soft drinks brands that sell
mostly offline?

Solution

Collect data on online conversations and

determine when a brand is mentioned
Link conversation data to offline sales data
period) and account for the effect of advert
(paid media)
Use statistical models to determine how dem
is stimulated by conversations
Simulate the value of online conversations us
the estimated model
18 brands, 19 months, 12 US markets
24

Average Monthly
Conversations Mentioning a Brand
(in 000s)

Revenue Change Due to 1,000

Additional Conversations Across All
Brands
(in US Dollars)
40
35
30
25
20
15
10
5
0

Facebook

Blog/Discussion
Boards

Twitter

Facebook

Blog/Discussion
Boards

Twitter

Conversations (in US Dollars)

60

FOR ALL BRANDS BI

AND SMAL

50
40
30
20
10
0

Facebook

Blog/Discussion Boards

Twitter

BIG DATA ANALYTICS

CAN INDEED HELP
BRAND
ENTERTAINMENT

Remember
Start simple and reuse existing tools and analyses (8
of the gains from the initial 20% of effort)
Think of the relevant unit of analysis (more than big d
it is intelligent data!)
Combine methods

How can we benefit?

Discovery and Sensing
Measurement and Prediction
Targeting and Personalization
Consider variety of data
Be creative in data structures and
analyses

THANK YOU

Market
Disruption

New York
Public
Library
Flash Mob
Partners &
Spade

September
2011

WHAT WORKS?
THE CASE OF
WARBY PARKER

Lets do some analysis with Google trends over time

the company has grown in terms of interest in the
market. We notice some significant spikes

Warby Parker
Searches

itor Traffic to Website

urce: Annual Report 2011

KEY
INTEGRATI
Google Trends: Searches
N AN
for Warby Parker in 2011 INTERACTI
N O
DIFFEREN
CHANNEL

NEW VERSUS
TRADITIONAL

Internet is growing fast and new media is gaining

importance
Traditional media is still huge Opportunities exist to
take advantage of their integration and interaction

ALL
AGES
EVEN
NEW
GES,
EVER
SEEN
FORE
)

TRAINING
IMAGES

Decompose the images in

different layers (using fil

Create a Dictionary of Vis

Words or Textons = Centroi
pixel clusters

Even if the system has ne

seen an image before, it
process it using the filters
then assign each pixel to
closest cluster and determ
how many of the Textons (v
words) we have in the ima

Final result is a distributio

Textons for each picture

WHY?

APPLICATION
SOCIAL MEDIA AND ONLINE
CONTENT CONSUMPTION:
PREDICTION AND SOCIAL
INFLUENCE

PREDICTING WITH SOCIAL

MEDIA

News Sites and

Facebook

100%#

Predicting
site visit

90%#
80%#
70%#
60%#

45#degree#line#

50%#
40%#

Full#Facebook#

30%#

Base#Model#

20%#

No#Facebook##

10%#
0%#

33% improvement in
Page Views prediction
over a model that was
performing already
extremely well
Improvement came
mostly from the
information of what
Facebook friends were
viewing at the news site

Own#Fcebook#
0%# 10%# 20%# 30%# 40%# 50%# 60%# 70%# 80%# 90%# 100%#
36