Title: Consumer perception and reviews on mobile phones:
An analysis using sentiment models in machine learning
Primary Supervisor: Prof. Dr. rer. nat. Thomas Wenger
Secondary Supervisor: Prof. Dr. rer. nat. Tobias Hagen
MBA Program Director: Prof. Dr. rer. pol. Rainer Fischer
Date of submission: 15th January, 2020
Date of presentation: 22nd January, 2020
Prior to the MBA (IBC) I have worked for more than 7 years at various
positions of the web and mobile application development spectrum in Japan,
South Korea, US, UK, India, Iceland and Germany.
Have been a software engineer by profession, and am a hobbyist musician.
Recently concluded an internship at Accenture in Burghausen, Germany.
Find it interesting to comprehend how people perceive products and
services, and keenly follow the development of smartphones.
Fascinated by craftsmanship behind the process of deriving actionable
insights from huge data-sets, which can help businesses in a huge way.
Mobile phone industry
one of the fastest growing sectors
defines success of consumer electronics firms today
global smartphone sales revenue: 522 billion USD
1.56 billion units being sold each year
Constant existential threat to players in the market due to innovations as
well as new entrants.
Amazon – one of the biggest online marketplaces, only matched by Alibaba.
Why Online Reviews?
Significance of online reviews:
90% of consumers read online reviews before visiting a business.
Online reviews have been shown to impact 67.7% of purchasing decisions.
84% of people trust online reviews as much as a personal recommendation.
Businesses risk losing 22% of business when potential customers find one negative article
on the first page of their search results and this risk grows to 44% and to almost 60%
with two and three negative articles respectively.
Why sentiment analysis?
Scalar ratings (typically 1-5) are not very helpful as:
The “why” for that rating or metric like average rating can’t be determined.
Numeric ratings are not comparable across segments and devices.
Unstructured raw data
Extracting human sentiments from written text
Sentiment analysis as a classification problem
Qualitative sentiment analysis
Overall aim: gaining actionable insights from customers’ voices
Polarity: discrete or continuous
Subjective and Objective sentiment analysis
ML-based models and lexicon-based VADER
Gathered by PromptCloud Web Scraping Service
Long-term data until 2018 made available under Creative Commons
license with all copyrights waived off.
Recent reviews data including reviews from mid 2018 to July 2019
purchased from PromptCloud for this research.
Purchased from Data Stock shop via:
After selection and pre-processing phases of the pipeline:
99708 long-term reviews and 49484 recent reviews were retained.
after de-duping, brand name harmonisation etc.
Datasets intentionally not unified to perform separate analyses.
Exploratory statistical analysis
Comparing performance of models for sentiment classification:
Support Vector Machine (with linear kernel)
naïve Bayes (Gaussian)
Random Forest and Ensemble Methods
Compound sentiment analysis using VADER and qualitative analysis on specific
Business use-cases and interpretation of findings.
Data selection and sanitisation
Exploratory Statistical Analysis
Counts, mean values, distribution of ratings among the reviews,
correlation between review length and perceived helpfulness
Compound Sentiment Analysis
Qualitative Sentiment Analysis
Data Encoding & Splitting
no ordinal relationship exists in unstructured textual data
binary values are preferred over integer encoding
typically used for normalising a set of labels, or for transforming non-
numerical labels to numerical ones, our use case: KNN
training, testing, hold-out validation
Business Implications of metrics
True negatives have far less business costs compared to false negatives
and false positives.
Example: missing a very receptive market for social video game in Taiwan,
based on tests using English localisation, leading to false negative!
Model Evaluation Metrics
percentage of correct predictions among total predictions.
when the model predicts positive, how often is it correct.
when the outcome is positive, how often is our model saying so.
harmonic mean of precision and recall.
better measure to seek a balance, based on business costs.
Insights leading to most positives recently on Android phones:
wireless charging, image stabilisation, curved edge
quad-core, heart rate, super AMOLED, battery life, snapdragon
Insights leading to most negatives recently on Android phones:
unlocking, Bixby, phone heat, Android crash
Pants pocket, bloatware apps, useless features
Phone perception linking to general brand image:
many highly positive iPhone reviews refer to MacBook Pro, Air, iPad Pro
post-sale customer service also seems to impact product perception
trends from rolling and expanding means on time series coincide with events
Descriptive, predictive and prescriptive analytics
data-driven decision making
uses abound in industries from video games, stock markets to medicine
Examples from currently thriving start-ups:
Gavagai – instant operational insights
Talkwalker – empowering brands socially
Aspectiva – acquired by Walmart for recommendation engine
Smartmunk – improving customer loyalty
Revuze – text mining on call center feedbacks, online CX, social media etc.