INDUSTRIAL TRAINING
Submitted in partial fulfillment of the
Requirements for the award of the degree
of
Bachelor of Technology
in
Artificial Intelligence & Data Science
By:
Vishakha Gupta (05713211921/AIDS/2025)
Department of Artificial Intelligence & Data Science
Guru Tegh Bahadur Institute of Technology
Guru Gobind Singh Indraprastha University
Dwarka, New Delhi
Year 2021-2025
Sentiment Analysis & Customer
Behavior Prediction for British Airways
Duration
1st July, 2024 – 31th July, 2024
By:
Vishakha Gupta (05713211921/AIDS/2025)
At
British Airways,
Forage
i
DECLARATION
I, Vishakha Gupta, hereby declare that all the work presented in this Summer
Training Report at British Airways, Forage for the partial fulfillment of the
requirements for the award of the degree of Bachelor of Technology in Artificial
Intelligence and Data Science, Guru Tegh Bahadur Institute of Technology,
affiliated to Guru Gobind Singh Indraprastha University Delhi is an authentic record
of our own work carried out at Forage, Delhi from 1st July, 2024 to 31st July, 2024.
Date:
(57/AIDS/2025) Vishakha Gupta
ii
CERTIFICATE
iii
ACKNOWLEDGEMENT
I would like to express my great gratitude towards Ms. Basanti Pal Nandi who has
given me support and suggestions. Without their help I could not have presented this
work upto the present standard. I also take this opportunity to give thanks to all others
who gave me support for the project or in other aspects of our study at Guru Tegh
Bahadur Institute of Technology .
Vishakha Gupta (57/AIDS/2025) Date:
Email: vishakharigel11@[Link]
iv
ABSTRACT
In today’s competitive airline industry, providing a great customer experience is
crucial for building lasting relationships and ensuring customer loyalty. British
Airways (BA), a major player in global air travel, wants to better understand how
passengers feel about its services to improve both efficiency and overall experience.
To do this, BA plans to gather and analyze feedback from passengers and internal
performance data to gain insights into every stage of the customer journey.
The first step is to collect data from sources like customer surveys, reviews, social
media, and operational performance reports. Sentiment analysis will be used to
evaluate what passengers are saying about key aspects of their experience, such as
booking, boarding, and inflight services. By combining this with demographic
information, flight routes, and service ratings, BA will segment its customer base and
use machine learning to identify patterns. This will help reveal areas where
improvements are needed.
Next, BA will focus on operational efficiency. By linking customer feedback with
operational metrics like flight delays, baggage handling, and check-in times, BA can
spot inefficiencies in its processes. Predictive models will help forecast potential
customer dissatisfaction, enabling BA to take proactive steps to address issues before
they impact passengers.
This analysis will give BA actionable insights to improve customer satisfaction and
streamline operations.
v
LIST OF FIGURES AND TABLES
Fig No Figure Name Page
1 Reviews from Skytrax Website 13
2 Mutual Information Analysis 15
3 PowerPoint Presentation on Task 1 27
4 PowerPoint Presentation on Task 2 27
vi
CONTENTS
Chapter Page No.
Title Page i
Declaration ii
Certificate iii
Acknowledgement iv
Abstract v
Tables and figures vi
1. Internship Overview 8-10
2.1 Introduction 9
2.2 Learning Objectives 9
2.3 Key Tasks and Deliverables 9
2.4 Learning Experience 10
2.5 Relevance of Internship to British Airways’ Business Strategy 10
2. Project Methodology 11-16
2.1. Introduction 12
2.2. Tasks Performed 12
2.2.1. Task 1: Web Scraping and Sentiment Analysis 12-14
2.2.2. Task 2: Classification Model for Customer Buying Behaviour 14-16
3. Requirement Analysis with SRS 17-19
3.1. Functional Requirements 18
3.2. Non-Functional Requirements 18
3.3. Software Requirement Specification (SRS) 19
4. Results and Findings 20-23
4.1. Task 1: Web Scraping and Sentiment Analysis 21-22
4.2 Task 2: Classification Model for Customer Buying Behaviour 22-23
5. Conclusion 24-25
6. References 25-26
7. Appendix 27-34
vii
CHAPTER -1
Internship Overview
8
2.1 Introduction
The British Airways Data Science Internship offered by Forage focused on solving
real-world challenges faced by one of the largest airlines globally. The internship
aimed to immerse participants in data science techniques that could improve
operational efficiency, enhance customer experiences, and optimize business
strategies. As a data science enthusiast, I was intrigued by the opportunity to work on
an internship aligned with a global airline.
2.2 Learning Objectives
The internship was designed to achieve several key learning objectives, blending
technical skills with business insights. It emphasized hands-on data analysis to uncover
patterns and trends in real-world datasets using Python and its libraries, enhancing my
data manipulation and analysis skills. The internship also aimed to improve my
programming proficiency, particularly in Python, within data science workflows.
2.3 Key Tasks and Deliverables
The scope of the British Airways Data Science Internship encompassed two primary
tasks, each designed to address key aspects of customer insights and behavior.
Task 1: Web Scraping and Sentiment Analysis
• Data Collection through Web Scraping: Web scraping was used to gather
customer reviews from the Skytrax website using tools like BeautifulSoup.
• Sentiment Analysis: NLP techniques and sentiment analysis libraries were
applied to classify reviews as positive or negative to assess customer
satisfaction.
• Visualization and Presentation: Insights from the analysis were visualized
with key trends presented in a PowerPoint for actionable recommendations.
Task 2: Classification Model for Customer Buying Behavior
• Creating a Classification Model: Developed a model using historical
customer transaction data to predict future buying behavior, applying
algorithms like logistic regression and decision trees.
• Visualization and Presentation: Analyzed and visualized the model’s
performance, and presented the findings in a PowerPoint to demonstrate its
9
predictive capabilities and provide recommendations for improving marketing
strategies and customer engagement.
2.4 Learning Experience
Throughout the internship, I significantly improved my technical and soft skills,
particularly in the following areas:
• Programming in Python: I gained proficiency in Python through extensive
use for data cleaning, processing, and machine learning model implementation.
• Data Visualization: I developed skills in creating clear and effective
visualizations using Matplotlib and Seaborn to present data insights.
• Statistical Analysis and Machine Learning: I applied statistical techniques
and basic machine learning algorithms, to solve business problems..
• Problem-Solving: Troubleshooting data analysis issues enhanced my
problem-solving skills and approach to debugging.
• Communication and Report Writing: I honed my ability to present complex
data findings clearly through written reports and presentations.
2.5 Relevance of Internship to British Airways’ Business Strategy
The data science insights generated during the internship helped British Airways
enhance customer service. For example, by analyzing customer behavior, the airline
could offer personalized travel experiences, optimize ticket prices, and improve in-
flight services. The internship tasks aligned with British Airways' goal to increase
operational efficiency. By analyzing flight performance data, I was able to provide
insights on improving flight schedules, reducing delays, and cutting operational costs.
The learnings from this internship could be applied to future data science projects
within British Airways. As the airline continued to embrace data-driven decision-
making, there would be opportunities to develop more sophisticated models for
predictive maintenance, customer satisfaction.
10
CHAPTER - 2
Project Methodology
11
2.1 Introduction
In today’s highly competitive aviation industry, understanding customer sentiment and
predicting buying behavior are crucial for enhancing service quality and driving
business success. During my data science internship with British Airways Forage, I
was tasked with two key objectives that aimed to leverage data-driven insights for
strategic improvements.
The first task involved applying web scraping techniques to gather unstructured data
from customer reviews on the Skytrax website. This data was then subjected to
sentiment analysis to classify reviews into positive and negative categories. By
analyzing customer feedback, we aimed to uncover underlying sentiments and trends,
providing valuable insights into customer perceptions of British Airways.
The second task focused on creating a classification model to predict customer buying
behavior. Using advanced machine learning techniques, we sought to develop a model
capable of forecasting purchasing patterns based on historical data.
This methodology chapter provides a comprehensive overview of the methods and
techniques used in these tasks. It details the processes that generated actionable
insights and strategic recommendations.
2.2 Tasks Performed
During my data science internship with British Airways Forage, I was tasked with two
key objectives that aimed to leverage data-driven insights for strategic improvements.
2.2.1 Task 1: Web Scraping and Sentiment Analysis
Data Collection
The website ([Link] is a
platform dedicated to collecting and displaying airline customer reviews. It provides a
space for passengers to share their experiences with airlines, including British
Airways, by offering detailed feedback about various aspects of the service such as
staff behavior, seat comfort, in-flight services, and overall satisfaction.
12
(Fig 2. Reviews from Skytrax Website)
Web Scraping
For web scraping, I employed tools such as BeautifulSoup (for parsing HTML) and
Requests (for making HTTP requests to access the website's content). I chose
BeautifulSoup due to its simplicity and effectiveness in handling large amounts of
HTML data. The scraping process involved identifying the relevant sections of the
website’s structure, such as the review text, review date, and customer ratings, by
analyzing the HTML tags and class attributes. I used Python for writing the scripts,
automating the process of collecting customer reviews, while adhering to ethical
scraping practices, ensuring I respected the website’s terms of use and implemented
delays to avoid overloading the server.
The data collected included customer-written feedback describing their experiences,
customer ratings on different aspects of the service, such as cabin cleanliness, seat
comfort, and staff friendliness, the date on which the review was posted, which helped
in tracking trends over time. This data was then stored in a structured format
(CSV/JSON) for further analysis.
Sentiment Analysis
Before performing sentiment analysis, the review text had to be cleaned and prepared.
1. Text Cleaning: This involves removing unnecessary elements like HTML
tags, special characters, and punctuation. The text is also converted to
lowercase for consistency.
13
2. Tokenization: The text is broken down into smaller pieces, typically
individual words (word tokenization), to analyze each word's contribution to
the sentiment.
3. POS Tagging: Parts of Speech (POS) tagging assigns grammatical labels
(noun, verb, etc.) to each word, aiding in tasks like lemmatization by preserving
word context.
4. Stopwords Removal: Common words (e.g., "the," "is") that add little meaning
are removed, reducing noise and focusing on important words.
5. Stemming and Lemmatization: Stemming chops off word endings, while
lemmatization uses POS tags for meaningful root forms.
Tools: The Natural Language Toolkit (NLTK) in Python supports tasks like
tokenization, POS tagging, stopword removal, stemming, and lemmatization.
For sentiment analysis, I utilized the VADER (Valence Aware Dictionary and
Sentiment Reasoner) library, a lexicon and rule-based sentiment analysis tool
designed for social media, but also effective in analyzing customer reviews. VADER
was chosen because it could effectively identify the polarity of the reviews and assign
a sentiment score, categorizing each review as positive, negative, or neutral.
Classification: Reviews were classified based on sentiment scores generated by
VADER and the machine learning models. I used thresholds to categorize the reviews:
• Positive Reviews: Sentiment scores above a defined threshold.
• Negative Reviews: Sentiment scores below a threshold.
• Neutral Reviews: Scores that fell within a middle range.
2.2.2 Task 2: Classification Model for Customer Buying Behavior
This task aimed to develop a machine learning model to predict whether a customer
will complete their booking based on a variety of features. The goal was to create an
accurate and reliable model, evaluated through metrics such as accuracy and AUC
scores, using algorithms like Random Forest and XGBoost classifiers.
Data Understanding
The dataset used in this project consists of 50,000 records, with 14 features capturing
various details of customer bookings, such as the number of passengers, sales channel,
14
trip type, booking origin, and more. The target variable in this dataset is
booking_complete, which indicates whether the booking was completed (1) or not (0).
The data contained no missing values, providing a clean basis for further analysis.
Data Preparation and Feature Selection
First, categorical variables such as route, trip_type, and sales_channel were converted
into numerical formats using label encoding and one-hot encoding. This
transformation ensured that these variables could be processed effectively by the
machine learning algorithms. Additionally, to ensure uniformity and improve model
performance, a MinMaxScaler was applied to scale the numerical features, bringing
them into a consistent range between 0 and 1.
To identify which features contributed the most to predicting booking completion, we
conducted a Mutual Information (MI) analysis. As a result, the top features identified
were route, booking_origin, flight_duration, length_of_stay, and
wants_extra_baggage. These features were then prioritized in the model training
process to test their impact on prediction accuracy.
(Fig 3. Mutual Information Analysis)
Train-Test Split: The data was split into training, validation, and test sets. We first
divided the data into 80% for training and 20% for testing, then further split the training
set into a 75-25% split for validation purposes. This process allowed us to train the
models on one portion of the data, validate performance on another portion, and finally
test the model's generalizability on unseen data.
15
Model Selection and Training
Several models were tested during this phase, with each model trained and evaluated
using different subsets of the data. The following models were used:
Model 1: Random Forest Classifier (Top 6 Features): The first model used the top
six features identified during feature selection. This model achieved an accuracy of
83.36% and an AUC score of 0.5657, demonstrating decent performance but with
room for improvement.
Model 2: Random Forest Classifier (All Features): The second model utilized all
available features, rather than focusing solely on the top six. This approach yielded a
higher accuracy of 84.76%, although the AUC score slightly decreased to 0.5479.
Despite this, the accuracy improvement made it a strong contender for the final model.
Model 3: XGBoost Classifier (Top 6 Features): Using the XGBoost classifier with
the top six features, this model achieved an accuracy of 84.87%. However, the AUC
score of 0.5005 suggested that the model struggled with distinguishing between the
two classes effectively, making it less suitable for adoption.
Model 4: XGBoost Classifier (All Features): The final model involved applying
XGBoost to all features. While the accuracy was comparable to the Random Forest
model, at 84.83%, the AUC score of 0.5065 indicated a limited ability to separate the
two classes.
Model Evaluation
After evaluating all models, the Random Forest Classifier using all features (Model
2) was selected as the final model for deployment. This decision was based on the
model's strong accuracy of 84.76% and reasonable AUC score of 0.5479, which
outperformed the other models tested. The slight increase in accuracy compared to the
other models made it the optimal choice for predicting customer booking completion.
Model Validation
To validate the chosen model's performance, it was tested on the reserved test dataset.
The model's final accuracy on this unseen data was 85.09%, with an AUC score of
0.5578. These results confirmed that the Random Forest model generalized well to
new data, supporting the decision to finalize and deploy it.
16
CHAPTER - 3
Requirement Analysis
With SRS
17
3.1 Functional Requirements
Functional requirements define the specific behavior or functions of the system. Based
on the project goals outlined in the internship report, the functional requirements
include:
• Web Scraping: The system must scrape airline reviews from designated
websites like Skytrax to collect unstructured customer feedback.
• Sentiment Analysis: The system should classify customer feedback into
positive, negative, or neutral categories using natural language processing
(NLP) tools such as VADER.
• Customer Behavior Prediction: The system must predict customer buying
behavior using machine learning techniques, like classification models
(Random Forest, XGBoost).
• Data Preprocessing: The system must handle data cleaning, transformation
(encoding categorical variables), and preparation for analysis.
• Model Evaluation: The system must validate machine learning models based
on accuracy and AUC score.
• Visualization: The system should generate clear and actionable data
visualizations for presenting results to stakeholders.
• User Interface: The solution needs to support data input, such as uploading
datasets, and display analysis results in a report or dashboard format.
3.2 Non-Functional Requirements
These define the performance characteristics and operational constraints of the system:
• Performance: The system should perform sentiment analysis and model
training within an acceptable timeframe, even with large datasets.
• Scalability: The system must be able to handle increased data volume and
additional features without significant performance degradation.
• Security: Data privacy must be maintained when handling sensitive customer
information, ensuring secure web scraping and storage.
• Reliability: The system should be robust enough to consistently perform data
scraping and analysis tasks without frequent failure.
• Usability: The user interface should be intuitive for users to upload data and
view results without requiring in-depth technical knowledge.
18
• Maintainability: The codebase should be structured and documented to allow
for easy updates and troubleshooting.
3.3 Software Requirement Specification (SRS)
The SRS defines the technical and software requirements needed to implement the
system:
• Programming Language: Python for web scraping (using BeautifulSoup),
data analysis, and machine learning.
• Frameworks and Libraries:
o Web Scraping: BeautifulSoup, Requests.
o NLP and Sentiment Analysis: Natural Language Toolkit (NLTK),
VADER Sentiment Analysis.
o Machine Learning: Scikit-learn for model building, XGBoost for
boosting models.
o Data Visualization: Matplotlib, Seaborn for plotting sentiment
analysis and model performance.
• Database: CSV or JSON file format for storing scraped data.
• Development Tools: Jupyter Notebook or any Python IDE for development.
• Operating System: The system should be platform-independent, functioning
on Windows, Linux, and macOS.
• Hardware Requirements: A typical development machine with at least 8GB
RAM for running machine learning models efficiently.
• External APIs: API support for integrating additional data sources or serving
predictive model results.
This requirement analysis defines the scope and technical expectations for the
development of the system described in the document.
19
CHAPTER - 4
Results and Findings
20
This chapter presents the key results and findings from the two primary tasks
undertaken during the data analysis for British Airways. Both tasks contribute valuable
insights into customer preferences, operational strengths, and areas for improvement,
ultimately aiding British Airways in enhancing their customer experience and business
strategies.
4.1 Task 1: Web Scraping and Sentiment Analysis
Sentiment Distribution and Analysis
The sentiment analysis of 1,000 customer reviews for British Airways revealed the
following distribution:
• Positive Reviews: 53.8% of the reviews were classified as positive, reflecting
general satisfaction among a majority of customers.
• Negative Reviews: 34.7% of the reviews were classified as negative,
highlighting specific pain points and areas of dissatisfaction.
• Neutral Reviews: 11.5% of the reviews were classified as neutral, showing a
balanced sentiment with neither strong satisfaction nor dissatisfaction.
This sentiment distribution indicates that while most customers had a favorable
experience with British Airways, a significant portion still encountered issues that need
to be addressed.
Sentiment Over Time
The sentiment analysis showed fluctuations in customer satisfaction over time. Periods
of increased negative sentiment were often correlated with operational issues, such as
peak travel seasons, strikes, or technical disruptions.
Key Findings and Insights
• Customer Satisfaction: The majority of customers had a positive experience,
especially praising customer service and flight comfort. However, the
significant number of negative reviews suggests operational issues like delays
and baggage mishandling are recurring pain points.
• Operational Improvements: British Airways could enhance customer
satisfaction by addressing frequent complaints related to delays and baggage
handling.
21
• Targeted Enhancements: Faster and more efficient customer support during
flight disruptions and improvements in baggage handling could help mitigate
some of the negative sentiment and elevate the overall customer experience.
• Word Cloud: A word cloud of the most common keywords found in the
reviews revealed prominent terms such as "Heathrow," "British Airways,"
"flight," and "London." This suggests that many reviews focused on flight
experiences, specific airports, and customer interactions with the airline.
4.2 Task 2: Classification Model for Customer Buying Behavior
The classification model aimed to predict customer booking completion based on a
variety of factors, including flight routes, demographics, travel preferences, and
booking behaviors. The Random Forest and XGBoost models were employed, with a
detailed analysis of the accuracy, AUC scores, and other performance metrics.
1. Model Performance
• Random Forest (with all features): Achieved an accuracy of 84.76% on the
validation set with an AUC score of 0.547, indicating strong overall
performance in predicting customer buying behavior.
• XGBoost (with all features): This model yielded an accuracy of 84.83% and
an AUC score of 0.506, which, while comparable, did not outperform the
Random Forest model.
• Final Model: The Random Forest model was selected as the final model due
to its superior accuracy of 85.09% and AUC score of 0.558 (indicating room
for improvement in handling true positive/false positive trade-offs), as well as
its ability to handle complex interactions between features.
2. Feature Importance
Mutual Information analysis identified the most important features influencing
customer booking completion:
• Route: The flight route was the most significant feature, indicating that the
specific route chosen by customers plays a critical role in predicting whether
they will complete a booking.
• Booking Origin: The customer's country or region of origin also had a
significant impact on the likelihood of completing a booking.
22
• Flight Duration: Longer flight durations correlated with higher chances of
booking completion, possibly indicating that customers tend to plan more
thoroughly for long journeys.
• Extra Baggage: Customers who opted for extra baggage were more likely to
complete their bookings, potentially reflecting a higher level of commitment to
their travel plans.
• Length of Stay: The length of stay was another important feature, with longer
stays indicating a higher likelihood of booking completion.
3. Insights
• Frequent Flyers: Customers with a history of frequent travel, especially those
enrolled in loyalty programs, were more likely to complete bookings.
• Early Bookings: Early bookings had a higher completion rate, suggesting that
customers who plan their trips well in advance are more committed to their
bookings.
• Preferred Routes and Destinations: Customers traveling on specific,
preferred routes showed a higher likelihood of booking completion, potentially
indicating destination loyalty.
These insights can help British Airways tailor their marketing strategies, focusing on
loyal customers, promoting early booking incentives, and optimizing offerings on key
routes to boost booking completions.
23
CHAPTER - 5
Conclusion
24
This project aimed to provide British Airways with data-driven insights into customer
behavior through two key tasks: sentiment analysis of airline reviews and a predictive
model for customer buying behavior.
Task 1: Web Scraping and Sentiment Analysis
Using web scraping and natural language processing techniques, we successfully
gathered and analyzed customer reviews from various sources. Sentiment analysis
revealed the overall customer satisfaction levels, classifying the feedback into positive
and negative categories. The analysis highlighted key pain points such as baggage
handling, flight delays, and in-flight services, which were often the source of negative
reviews. On the other hand, positive reviews frequently praised the airline’s customer
service, comfort, and efficient booking experience. These findings offer actionable
insights for British Airways to improve their services by focusing on areas that directly
affect customer satisfaction.
Task 2: Predictive Model for Customer Buying Behavior
In the second task, a Random Forest classifier was developed to predict whether
customers would complete their booking based on various features such as flight route,
booking origin, and travel preferences. The model achieved an accuracy of 85.09%,
with route, booking origin, and flight duration emerging as the top factors influencing
buying behavior. Although the model performed well in terms of accuracy, the AUC
score of 0.558 suggests that further tuning is needed to better balance precision and
recall. These findings offer valuable insights into which customer segments are more
likely to purchase, enabling British Airways to target these customers with
personalized marketing strategies and improve conversion rates.
Together, both tasks provide British Airways with a comprehensive understanding of
customer sentiments and behavior. By leveraging these insights, the airline can
enhance customer satisfaction, optimize marketing strategies, and improve overall
business performance.
25
REFERENCES
1. [Link]
2. [Link]
3. [Link]
vader-works-716d272acac0
4. [Link]
5. [Link]
[Link]/stable/modules/generated/sklearn.feature_selection.mutual_info_cla
[Link]
6. [Link]
7. [Link]
8. [Link]
9. [Link]
10. [Link]
11. [Link]
[Link]
12. [Link]
course/classification/precision-recall
13. [Link]
14. [Link]
15. [Link]
16. [Link]
17. [Link]
18. [Link]
19. [Link]
20. [Link]
21. [Link]
26
APPENDIX A
SCREEN SHOTS
Task 1 : Web Scraping and Sentiment Analysis
(Fig 4. PowerPoint Presentation on Task 1)
Task 2: Predictive Model for Customer Buying Behavior
(Fig 5. PowerPoint Presentation on Task 2)
27
APPENDIX B
SOURCE CODE
Task 1 : Web Scraping and Sentiment Analysis
28
29
30
31
Task 2: Predictive Model for Customer Buying Behavior
32
33
34
