The document discusses developing a sentiment analysis model to classify movie reviews as positive or negative sentiment in order to better recommend movies to users. It evaluates several classification algorithms and finds that a Naive Bayes model achieves the highest accuracy of 78.46%. Additionally, it outlines how improving the accuracy of sentiment analysis could benefit applications like reducing movie rating costs, increasing revenues from movie sequels, expanding the lifetime value of subscribers, and generating more ad revenue from recommendation websites.

2. So tell me…
Which
movie
should I
watch?
Opinion Extraction
Better Promotion
Quick Review
classification
Rich User Engagement
Personalization
Accurate Rating
Good Recommendation
Minimum Search

3. CLASSIFICATION
 Training Dataset has 8000 reviews
of different movies
 Given Variables: Phrase Id,
Sentence Id, Phrase, Sentiment
 Output Variable: Sentiment (0 or
1)
 Input Variables: Features extracted
through Lightside tool
Confusion matrix
Act Pred 0 1
0 2501 771
1 744 2863
Confusion matrix
Act Pred 0 1
0 2598 680
1 803 2804
Confusion matrix
Act Pred 0 1
0 2435 843
1 900 2707
Logistic Regression Naïve Bayes
Support Vector
Machines
 Every review was initially broken into phrases in
separate rows with different phrase IDs
 We only took the phrase containing the entire review
and discarded the other phrases
 Initially 5 sentiments in training dataset: +ve,
Somewhat +ve, Neutral, -ve and Somewhat –ve
 Discarded the Neutral sentiments and grouped:
i) +ve and Somewhat +ve into 1
ii) -ve and Somewhat –ve into 0
Accuracy: 78% Accuracy: 78.46%
Accuracy:
74.68%%
Models Used
Highest accuracy was achieved in Naïve Bayes with 21.54% error
Variables/Tool Used Cleaning Steps

4. Can we do better? ………… YES!
Eliminate need of manual ratings
1000 reviews per movie
20 movies in a week
Whopping $ 3000 per movie!! Just to rate
them!
$3000/wk * 50wks/yr=
$1,50,000/ year
Improve the success of Sequels
70% success rate for a blockbuster’s sequel
20 sequels on an average in a year
Expected Revenue from sequel = $ 1400
million/year
At mere 1% of
revenues as
commission for
improvement(90%)=
$200 mn*1% = $2 mn
Expanding the subscriber’s Lifetime value:
Present user churn rate @ 50% accuracy= 25%
LTV = ARPA x gross profit margin / customer churn
Average LTV = $8x11%/25% = $ 3.2/user
Increase LTV by 7
times for an rating
improvement from 50%
to 78%
Ad revenue on recommendation websites
10 million unique users monthly
Average revenue per engagement = $ 0.5
Current Revenue from 50% accuracy = $ 60 million
Potential Savings at 78%
accuracy = $1.018
million

5. GUESS THE REVIEW SENTIMENT!!
Review Guess the
sentiment
Rotten
tomatoes’
Our prediction
The movie is so thoughtlessly assembled
Director Tom Dey demonstrated a knack for mixing action and
idiosyncratic humor in his charming 2000 debut Shanghai Noon
, but Showtime 's uninspired send-up of TV cop show cliches
mostly leaves him shooting blanks
Roman Polanski directs The Pianist like a surgeon mends a
broken heart; very meticulously but without any passion
` Synthetic ' is the best description of this well-meaning ,
beautifully produced film that sacrifices its promise for a high-
powered star pedigree
A film of empty , fetishistic violence in which murder is casual
and fun

6. WAY FORWARD
2) Category expansion:
Expose the sentiment analysis as API for
consumption by books, entertainment and
ecommerce websites
Suggest right movie for improving
TRP
No precise method behind screening
movies on TV network - Flat rate
based on popularity at box office
3) Algorithmic Improvement :
Improve Algorithms to interpret ambiguous
phrases
eg: This is great – this is not great - this could be
great - if this were great – this is just great
•Can we make it
more robust?
•Can we expand
the market scope?
•Can we reuse our
model?

7. Not sure which
movie to watch this
weekend?
You know who to
ask 