1. Marketing Strategy
7 drinks
Application of
Analytics & Business Intelligence
By:- (Group 7)
Ridhima Budhwar
Ryan Subhan
Michael Sawyer
James Ma
Steve Mendonca
Date: 05/04/2016
2. About the Company
Project Approach
Initial Analysis
Text Mining
Clustering
Topic modelling
Sentiment Analysis
Trending Tweets
Findings
Technical Challenge
References
Thank you!
OUTLINE
3. ABOUT THE COMPANY
7DRINKS - Decade old American multinational food and beverage corporation
HEADQUARTER - Newark, New Jersey, United States,
SERVICES - Manufacturing, marketing, and distribution of grain-based snacks and beverages.
$-
$500,000.00
$1,000,000.00
$1,500,000.00
$2,000,000.00
$2,500,000.00
$3,000,000.00
$3,500,000.00
2013 2014 2015
Company Performance
Cost Price Sales
GOALS:
Increase Sales
Market Penetration
22%
22%
11%
14%
4%
27%
Market share
Pepsi
Frito Lays
Tropicana
Gatorade
7drinks
Others
How could we do that?
Market Research
Market Strategy:
Ad campaigns
Brand Ambassador
Disclaimer:Dataonthisslideisfictitiousandmadeup.
4. PROJECT APPROACH
TEXT MINING:
Process of examining large collections of
unstructured textual resources
85% of all data available is unstructured
ABOUT DATA:
Data: Live twitter data
Number of Tweets: 1000
HIGH LEVEL STEPS:
Check what the competitors are doing?
Text mining for trending topic on Twitter
(Relevant) Famous celebrities
Example
COULD IT WORK??
5. INITIAL ANALYSIS
PROMOTION BRAINSTORMING:
Mine tweets of largest competitors for trending
topics, then brainstorm similar ideas for viral
marketing campaign(s)
Mine tweets in densely-populated areas for trending
celebrities, then seek to either sign them or their
local-market counterparts as spokesperson(s)
IMPLEMENTATION:
Most recent 200 tweets of five largest
competitors/brands = 1000 tweets
Five largest US cities: NYC, LA, Chicago,
Philadelphia, Houston
6. TEXT MINING
OUR PROCESS:
Acquire the tweets, assemble into a
single data frame
Convert to a corpus:
Change all cases to lower
Remove URLs
Remove unprintable
characters, like emoticons
Remove stopwords
Stem the corpus
Create a term-document matrix
and find most frequent terms
Reduce sparsity
7. INITIAL RESULTS
MOST COMMON TERMS:
“Thanks”, “sharing”, and “make”
Need to have a human actively managing the social
media interaction
Some branded terms, like “Shareacoke” and
“Pepsicola”, also had N >= 25
Terms that include the brand in hashtags are one
possible avenue
10. SENTIMENT ANALYSIS
SENTIMENT ANALYSIS:
Tries to search and group words
based on emotions:
Positive/negative
Anger/disgust/fear/
joy/sadness/surprise
Did not successfully distinguish
between any of the tweets – a
stemming issue?
11. TRENDING TWEETS
What is EVERYONE talking tweeting about?
AND WHERE???
TwitterR: Search keywords, Search UserTweets
getTrends()
Fetches latest Trends dependent on Location
WOEID = Where on earth Identifiers
Cities
States
Countries
Goal: Trend about 7drinks marketing!
“7Drinks”
12. TRENDING - RESULTS
Example: Sunday 5/01: Fetch latest Trends in top 5 Major Cities
New York
LA
Chicago
Houston
Philadelphia
Clayton Kershaw Matt Slauson Klay Thompson Monta Ellis
Next Steps:
Monitor Trends
Monitor # of Followers
Engage with Celebrities
Launch Marketing Campaign
Promote 7Drinks!
Results
13. FINDINGS
MARKETING RECOMMENDATIONS:
Consider a campaign to place the “7 Drinks” name in hashtag form, since those are the
majority of trending topics and easy to mine for associations
Certain celebrities do retweet and use hashtags more frequently than others; Selena Gomez
was one specific example
Sports figures on playoff teams are a common theme across major metropolitan areas (and
in the case of NYC, regardless of whether the player plays for a local team), so consider an
endorsement deal with player(s) whose image fits the brand’s, e.g. Klay Thompson
NEXT STEPS:
For more completeness, continue ideation by monitoring Twitter at regular intervals over a
longer period
Consider other channels to implement sentiment analysis, which could be a powerful tool
14. CHALLENGES
LIMITATIONS OF TWITTER MINING:
R tools to query tweets and related
metadata can be dimensionally limited:
Single username at a time
Favorites of that username
Single geocode
“Trending topics” accessible, but
decided by a black box
Hashtags are difficult to parse for
meaning, even more so than semantics in
NLP
Certain functionality may not always be
supported – the SENTIMENT package was
removed from CRAN
TIME LIMITATIONS:
Cap on number of tweets that can be
accessed – means actual time horizon of
analysis is usually very short
Therefore, would need to repeatedly
conduct these analyses over a longer
period in order to get actionable ideas
Data on followers needs another iteration
to drill down and get their followers –
computationally intensive
Accounts with large numbers of followers
take even longer to get metadata on
Once we’ve organized the cleaned words, we can then run some clustering techniques to make some associations: Hierarchical clustering helps find dependencies between words, which we visualized with this dendrogram. It seems like the LHS is related to a campaign run by Coca-Cola around sharing pictures on Twitter with certain hashtags, and having those serve as entries into a contest. Since one of their current promotions has been “Share a Coke,” this is one possible way of boosting the signal of that campaign. The RHS seems to be more the responses of whoever is running the Twitter accounts and encouraging users to “Like” the same pages on Facebook, so that may be one form of cross-channel marketing. Lastly, we used K-means clustering to find particular common phrases, which unearth similar conclusions.
Now we try to reclassify and quantify the topics in the text using topic modelling, which tries to guess which keywords are the topics, which words belong to each topic, and how frequently those topics appear. The picture sharing campaign seemed to get more traction earlier in the month of April, but towards the end, it was overtaken by one by Red Bull.
One last methodology we tried was sentiment analysis, which tries to search and group words based on the emotions associated with them, in order to numerically figure out how participants feel about certain topics. The tools available in R allow a spectrum of emotions, from anger, disgust and fear, to sadness, joy, and surprise. Also, another capability is distinguishing whether a tweet has a positive or negative connotation. While unfortunately we ran into a challenge in that each tweet seemed to be given the same rating, in theory what this could do is allow marketers to measure the response to their campaigns in real time, without manually sifting through tweets, and adjust accordingly.