See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/377222561

A Master's Thesis Report on I. Toxic Comment Classification using

Bidirectional LSTM and TensorFlow II. Data Visualization and Linear
Regression Analysis in Python. Python Programm...

Thesis · January 2024

DOI: 10.13140/RG.2.2.31067.34084

CITATIONS READS

0 612

1 author:

George Joseph
IUBH Berlin
1 PUBLICATION 0 CITATIONS

SEE PROFILE

All content following this page was uploaded by George Joseph on 08 January 2024.

The user has requested enhancement of the downloaded file.

 A Master’s Thesis Report on

I. Toxic Comment Classification using Bidirectional LSTM and TensorFlow

II. Data Visualization and Linear Regression Analysis in Python.

Python Programming Assignment - DLMDSPWP01

Course of Study: MS in Computer Science

Author: George Joseph
Matriculation Number: 9212153
Tutor: Dr. Cosmina Croitoru
Date: October 2023
 ACKNOWLEDGMENT

I would like to express my appreciation and thanks to Dr. Cosmina Croitoru, who has been a
supervisor and mentor, throughout this project. Her guidance, support and feedback have been
invaluable in helping me complete this thesis successfully. I also want to extend my gratitude to my
friends who have dedicated their time and effort to help and guide me through this project. Your
collaboration, enthusiasm and shared commitment to excellence have played a role in achieving
our goals. Working with such a committed team has been an enriching experience.

I want to emphasize that without the support, guidance and contributions of Dr. Croitoru and my
friends, this project would not have been possible. I am genuinely grateful for your assistance. I am
excited about the potential, for collaborations.

Thank you for being a part of this journey.

Sincerely
George Joseph

I
 ABSTRACT

The goal of this Python project is to use deep learning and Natural Language Processing (NLP)
techniques to create a toxicity detection system. Using a dataset of text comments and their
corresponding toxicity labels, the project trains a machine learning model that can recognize
different types of toxicity, including "toxic," "severe toxic," "obscene," "threat," "insult," and "identity
hate."

The project is clearly organized into discrete sections and follows the guidelines for academic
writing. The introduction provides an overview of the assignment's rationale, goal, and parameters.
The creation of the NLP model, including data preprocessing, model architecture, and training
protocols, is the primary emphasis of the project. The Bidirectional LSTM neural network that
powers the toxicity detection model can parse text sequences for classification. The project uses
TensorFlow and Keras for model development.

The main conclusions of the project are outlined in the conclusion, which also showcases the
model's effectiveness in the toxin detection assignment. Metrics for training and validation, like
accuracy and loss, are displayed and plotted. The predictions of the model are applied to text input
supplied by users, enabling toxicity evaluation in real-time.

Two separate files make up the Python project: one is used for training the model, and the other is
used to evaluate the model with user input. In order to evaluate text toxicity, the latter file uses a
pre-trained tokenizer and a saved model.

By showing how to create a useful and effective toxicity detection model, this effort makes a
significant contribution to the field of NLP. Its ramifications include managing online communities,
moderating content, and identifying offensive language.

The Python project places a strong emphasis on formal writing requirements for academic writing,
such as organization, clarity, and conformity to rules. It is evidence of the student's aptitude for
choosing a research topic, carrying out experiments, and preparing results for academic
presentations.

II
CONTENTS
ACKNOWLEDGMENT
ABSTRACT
LIST OF FIGURES
SECTION I. RESEARCH TOPIC
CHAPTER 1. INTRODUCTION
1.1. Background
1.2. Aim
1.3. Objective
1.4. Research Question
CHAPTER 2. RELATED WORK
CHAPTER 3. DATA AND METHODOLOGY
I. Training the Toxicity Detection Model
3.1.1. Data Preprocessing
3.1.2. Model Architecture
3.1.2. Model Architecture
II Evaluating User Input for Toxicity
3.2.1. Model Loading
3.2.2. User Input Evaluation
3.2.3. Interpretation of Results
CHAPTER 4. CONCLUSION
CHAPTER 5. FUTURE USAGE
SECTION II. WRITTEN ASSIGNMENT
1. INTRODUCTION
2. DATA ANALYSIS
3. CONCLUSION
REFERENCE
LIST OF APPENDICES
1. RESEARCH TOPIC CODE
2. WRITTEN ASSIGNMENT CODE
GITHUB REPOSITORY LINK

III
 LIST OF FIGURES

Fig 1. Shows the sample from the training dataset we use for this model
Fig.2 Visualisation of Training v/s Validation loss
Fig 3. Visualisation of Training vs Validation Accuracy
Fig 4. The result of the model training validation loss and accuracy at each epoch
Fig 5. : Result of user input data showing toxic language
Fig 6. The result of the user input data shows non-toxic and toxic languages.
Fig 7. Data plot of x against y1, y2, y3 and y4 values
Fig 8. Scatter plot of x against y1.
Fig 9. Scatter plot of x against y2.
Fig 10. Scatter plot of x against y3
Fig 11. Scatter plot of x against y4.
Fig 12. Readings of RMSE, MAE AND R-Squared (R2) Values.
Fig 13. Plot of Linear Regression Model with actual value

IV
 SECTION I: RESEARCH TOPIC

Toxic Comment Classification using Bidirectional LSTM and TensorFlow

1
 CHAPTER 1. INTRODUCTION

1.1. Background
Python is a highly adaptable and powerful programming language that is widely used in many
different fields, such as machine learning and natural language processing (NLP). Its simplicity and
readability make it a favourite among developers and data scientists alike. Python's rich ecosystem
of libraries and frameworks, such as TensorFlow and Keras, empowers us to harness the potential
of cutting-edge technologies. This project sets out to demonstrate the use of Python in a natural
language processing project, focusing on sentiment analysis and text classification. The
comprehensive libraries that Python offers for machine learning, data manipulation, and
visualization give programmers the means to create and implement robust natural language
processing (NLP) models.
Moreover, Python has developed into a strong tool in the artificial intelligence and natural language
processing domains for text analysis and sentiment classification. In order to address an important
and contemporary issue, this project takes advantage of Python's capabilities to recognize harmful
content and foul language in user-generated text. Automated content moderation is now necessary
in modern digital contexts due to concerns about the safety and quality of online debate sparked
by the extensive usage of online platforms. A detailed explanation of developing and utilizing a
deep learning-based toxicity detection model can be found in this thesis. This study addresses the
significance of automated toxicity detection in the digital age by classifying user-generated
language as toxic or non-toxic using a big dataset and bidirectional Long Short-Term Memory
(LSTM) networks.

1.2. Aim
Concerns over the quality and safety of online conversations have grown as a result of the
expansion of online platforms. In the digital age, offensive language, damaging user-generated
content, and toxic content have all become a significant issue. The goal of this research is to
develop an automated toxicity detection system in order to solve these problems. The aim of this
research revolves around the creation of a deep learning model using Python that can categorize
user-generated text as toxic or non-toxic. This system attempts to improve the general quality of
online discourse and offer a workable solution for content moderation by utilizing deep learning and
natural language processing techniques.

2
1.3. Objective
The primary objective of this project is to develop an automated content moderation system that
can accurately detect toxic language and harmful content in user-generated text. It aims to create a
deep learning-based model using Python, specifically utilizing bidirectional Long Short-Term
Memory (LSTM) networks. The project focuses on training and fine-tuning this model to achieve
high accuracy in toxicity detection.

1.4. Research Question

Can a deep learning-based toxicity detection model, built using Python and bidirectional LSTM
networks, effectively categorize user-generated text as toxic or non-toxic, and thereby contribute to
automated content moderation in digital platforms?

CHAPTER 2. RELATED WORK

In the realm of hate speech detection, significant research has been conducted since 2010.
Various studies have contributed to this field, exploring different approaches to identifying and
addressing hate speech and offensive content. Some notable research includes the work of Kwok
and Wang (2013), Burnap and Williams (2015), Djuric et al. (2015), Davidson et al. (2017),
Malmasi and Zampieri (2018), Schmidt and Wiegand (2017), Fortuna and Nunes (2018), ElSherief
et al. (2018), Gamb¨ack and Sikdar (2017), Zhang et al. (2018), and Mathur et al. (2018).

Schmidt and Wiegand (2017) and Fortuna and Nunes (2018) conducted reviews of hate speech
detection approaches, shedding light on the diverse strategies employed in this domain. Kwok and
Wang (2013) utilized machine learning methods, including bag-of-words and bi-gram features, to
classify tweets as either "racist" or "non-racist." Burnap and Williams (2015) developed a
supervised algorithm for detecting hateful and antagonistic content in Twitter using an ensemble of
classifiers. Djuric et al. (2015) used neural language models to learn low-dimensional
representations of social media comments for hate speech detection. Davidson et al. (2017)
employed n-gram features, TF-IDF scores, and crowd-sourced hate speech lexicons with various
classifiers to distinguish hate speech from other offensive language. Malmasi and Zampieri (2018)
utilized features such as n-grams, skip-grams, and clustering-based word representations in their
ensemble classifier for hate speech detection.

3
Furthermore, research has extended to aggression detection, focusing on various forms of
aggression, such as overt aggression, covert aggression, and non-aggression. Notable work in this
area includes studies by Aroyehun and Gelbukh (2018), Madisetty and Desarkar (2018), Raiyani et
al. (2018), and Kumar et al. (2018b). These studies employed techniques such as LSTM and CNN
to identify aggression in text. Importantly, research has also explored offensive language
identification in non-English languages, such as German, Hindi, Hinglish (Hindi-English), Slovene,
and Chinese. These studies, including work by Wiegand et al. (2018), Kumar et al. (2018b), Mathur
et al. (2018), Fiˇser et al. (2017), and Su et al. (2017), adapted and developed models to address
offensive language in these specific linguistic contexts.

Recent advancements in natural language processing (NLP) have led to the increased popularity
of tasks related to hate speech detection, toxicity assessment, and offensive language
identification. These tasks pose unique challenges, as offensive language is often implicit and
context-dependent. Traditional approaches based on lexical analysis and bag-of-words
representations have limitations in handling such cases. Deep learning methods, such as Long
Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNN), have shown
significant improvements in handling these tasks compared to earlier complex classifiers. While
deep learning models have shown promise, they demand substantial amounts of annotated
training data and are often tailored to specific tasks and languages. Transferring these models to
different languages or tasks can be problematic due to their specificity. In light of this, language
models like BERT have gained prominence. BERT, a bidirectional model, allows context to be
learned from both the left and right of words, enabling it to capture complex contextual information.
This related work chapter highlights the diversity of research in the field, ranging from hate speech
detection to aggression identification and offensive language identification in multiple languages. It
underscores the significance of deep learning models and their impact on addressing offensive
language and hate speech. Moreover, it emphasizes the growing importance of contextual
language models like BERT in tackling these challenges effectively. In the following sections, we
will delve into the specifics of our approach to addressing offensive language in German tweets by
leveraging BERT, providing insights into our unique contributions to this evolving research area.

4
 CHAPTER 3. DATA AND METHODOLOGY

I. Training the Toxicity Detection Model

3.1.1. Data Preprocessing

In this Python project, we perform essential data preprocessing stages to ready our dataset for
training a toxicity detection model using deep learning techniques. The dataset we utilize is
obtained from the Kaggle competition titled "Jigsaw Toxic Comment Classification Challenge." This
dataset consists of text comments that we aim to classify into various toxicity categories. The data
preprocessing pipeline includes the following key steps:

Data Loading: We initialize by loading the dataset from a specified local directory which was
downloaded from the Kaggle competition titled "Jigsaw Toxic Comment Classification Challenge.
This dataset comprises user-generated text comments (from almost 159570 users collected) from
online platforms. To efficiently manage and manipulate this data, we utilize popular libraries such
as Pandas for data loading. Pandas makes data organizing and interpretation easier, enabling us
to work with structured datasets with ease.

5
 Fig 1. Shows the sample from the training dataset we use for this model

Text Processing: After using Pandas to load the dataset, we start analyzing the text comments.
These remarks frequently contain textual elements like punctuation, special characters, and other
noises. We efficiently clean, format, and preprocess the text using Pandas and Python's string
manipulation features to guarantee the data's quality and get it ready for additional analysis.

Text Tokenization: To enable machine learning on textual data, we employ a tokenizer that
translates the text into numerical representations. With the help of Pandas and the joblib package,
we can effectively use the tokenizer with a 10,000-word vocabulary. This guarantees that the
10,000 words in the dataset that occur the most frequently are mapped to distinct number values.
This crucial stage, which prepares the groundwork for deep learning and allows the model to
operate with organized numerical data, is made easier by the combined use of joblib and pandas.

6
Sequence Padding: For text classification, consistent sequence lengths are critical to successful
model training. Text comments, however, can differ in length. We use joblib and Pandas to
regularly pad the tokenized sequences in order to solve this problem. For this project, we have
selected a sequence length of 200 words. In order to maintain homogeneity in the model's input
dimensions, padding may be necessary, such as by truncating text or adding special tokens.

The preparation of our data for later deep learning-based toxicity detection model training is largely
dependent on these preprocessing stages. They enable the model to effectively learn from and
anticipate textual material, and the choice of sequence length ensures that crucial information is
captured from every comment while maintaining computing efficiency. Pandas and joblib together
simplify these preparatory chores, allowing for easy dataset manipulation and transformation.

3.1.2. Model Architecture

Our Python project uses the Keras package to build a deep-learning model for toxicity identification
in user-generated text. An essential part of our work is the model architecture, which is made to
efficiently categorize text comments into various toxicity groups. The architecture of the model is as
follows:

Embedding Layer: We include an embedding layer at the beginning of our model. In order to
convert words into numerical representations, this layer is essential. In our project, we select an
output dimension of 64 and set the embedding layer's input dimension to 10,000, meaning it can
handle up to 10,000 unique words. Through this technique, the words in the comments are
transformed into dense vectors that the model can process more efficiently.

Bidirectional LSTM Layer: Following the embedding layer, we employ a Bidirectional Long
Short-Term Memory (LSTM) layer with 64 units. LSTM networks are a great option for applications
involving natural language processing since they can handle sequential input with ease. The
bidirectional feature improves the model's capacity to identify complexities and long-range
connections in the text input by allowing it to take into account the context of both past and future
words.

Dense Layer: After the LSTM layer, we include a dense layer comprising 64 units with Rectified
Linear Unit (ReLU) activation. The goal of the thick layer is to add non-linearity to the model so that
it can help extract more intricate patterns from the input. The ReLU activation function expedites
training and aids in mitigating vanishing gradient issues.

7
Dense Output Layer: Six units make up the dense output layer at the end of our model, one for
each toxicity category we want to categorize. In this layer, we apply a sigmoid activation function to
simplify multi-label categorization. With this option, our algorithm can handle many labels at once
and independently forecast if each toxicity category will be present or absent for a given text
remark.

Our toxicity detection model is based on the architecture presented here. Our model's ability to
comprehend and categorize text comments according to their level of toxicity is made possible by
the combination of layers; this helps to improve online safety and provides important insights for
content moderation.

3.1.3. Model Training

The Model Training stage is a critical component of the entire process. During this phase, we take
the preprocessed dataset, compile our deep learning model with specific configurations, and train it
to acquire the ability to classify text comments effectively.

Loss Function and Optimizer: The model is trained using a binary cross-entropy loss function
and the Adam optimizer. The multi-label classification task is a good fit for the binary cross-entropy
loss function. By quantifying the difference between the model's predicted and real labels, it helps
the model make parameter adjustments to reduce this difference. By optimizing the model's
parameters, the Adam optimizer—which is renowned for its effectiveness and adaptable learning
rate—increases the predicted accuracy of the model.

Visualization with Matplotlib: We use Matplotlib, a powerful visualization library, to obtain insights
into the model's training process. We can produce informative visuals with Matplotlib that show the
evolution of the model's performance over time. It is easy to see how the model improves accuracy
and minimizes loss as it learns from the training data by creating dynamic graphs that show the
training loss and accuracy. These visuals provide insightful information about the model's learning
process.
The graph figures of the evaluation is mentioned below:
1. Training v/s Validation loss

8
 Fig.2 Visualisation of Training v/s Validation loss
2. Training v/s Validation Accuracy

Fig 3. Visualisation of Training vs Validation Accuracy 9

Epochs: In the context of model training, "epochs" refer to the number of complete passes the
entire training dataset makes through the model. It is similar to iterating through the training
process several times, where the model refines its parameters at each epoch. Selecting the
appropriate number of epochs is crucial since an excessive number can cause overfitting and an
insufficient number can cause underfitting. We have selected five epochs for this research.
Keeping track of training and validation loss throughout epochs aids in finding the ideal equilibrium.

Making Predictions: Once the model is sufficiently trained, it's ready to make predictions on the
test data, consisting of text comments. These predictions do two things: first, they let us assess
how well the model performs; second, and perhaps more importantly, they make real-world
applications possible. To make sure the model is prepared for real-world settings, real-world data is
used to evaluate its capacity to classify comments for toxicity.

Model and Tokenizer Saving: To ensure the preservation and reusability of the trained model, we
save it as a file, commonly with the extension .h5 (Large volumes of data are frequently stored and
managed in an organized manner using.h5 files, which stand for Hierarchical Data Format version
5. The scientific and data analytic groups utilize it extensively due to its adaptability and
effectiveness in managing intricate data structures.). Because of this, we are able to effortlessly
load the learned model for usage in the future without having to retrain it. Furthermore, we store
the tokenizer that is utilized for text preprocessing, so that we can consistently prepare text data for
assessment across sessions and machines.
The results are provided below

10
 Fig 4. The result of the model training validation loss and accuracy at each epoch

This model training process is an integral part of our project, empowering our deep learning model
to become proficient at identifying toxic language within user-generated text. Using the Adam
optimizer, the binary cross-entropy loss function, and Matplotlib to visualize the training process,
we make sure that our model is always improving in terms of toxicity classification accuracy and
efficiency. Model and tokenizer saving facilitates model deployment across a range of applications,
from content moderation to real-time toxicity detection; the idea of "epochs" helps strike the correct
balance between overfitting and underfitting.

11
 II. Evaluating User Input for Toxicity

3.2.1. Model Loading

Here, using the keras.models.load_model() function from TensorFlow and Keras, we load the
pre-trained model stored in a file with the extension .h5, which was trained. This pre-trained model
has already absorbed valuable knowledge from extensive training data, making it a highly capable
toxicity classifier. By loading this model, we eliminate the need to train a new model from scratch
each time we want to evaluate user input. This results in faster and more efficient toxicity
assessments.
Additionally, We also load the saved tokenizer using the joblib.load() function to facilitate the
conversion of user input text into a format that the model can comprehend. The tokenizer ensures
that input data is processed consistently and in a manner consistent with the procedures during
model training. This step is essential for maintaining the accuracy of the toxicity evaluation.
We have enabled users to evaluate user-generated material in real-time without having to go
through the time-consuming process of model training by incorporating model loading into the
evaluation process. Because effective content filtering is made possible, this not only improves the
user experience but also promotes a safer and more civilized digital space.

3.2.2. User Input Evaluation

We offer an interface for users to assess text input in real-time in this Python toxicity detection
project. By recognizing potentially harmful content, this enables users to maintain a more secure
and civilized online space.

Preprocessing User Input: Users are prompted to enter text, which is then subjected to
preprocessing steps consistent with the techniques applied during model training. This process
involves tokenization and padding of the user's input to make it compatible with the model's
requirements.

Toxicity Detection: Using our previously trained toxicity detection model is the core of evaluating
user input. This model has been trained on a large amount of data, making it proficient in detecting
toxicity in text. It assesses user input for any indications of harmful language or toxicity.

Determination of Toxicity: Toxic language may be detected in user input by the script, which
goes one step further. Based on a predetermined toxicity threshold that can be altered to meet
particular needs, this decision is made. An indication that the user input contains harmful language
is displayed if the toxicity forecasts for the input surpass this cutoff.

12
3.2.3. Interpretation of Results
The evaluation process provides clear and user-friendly results to inform users about the nature of
the input they provided. Users are presented with the following information:

Toxicity Label: Users receive immediate feedback in the form of a toxicity label. This label makes
the assessment process easier to understand and helps users make well-informed decisions about
the content by indicating whether or not the input contains harmful words.

Optional Individual Toxicity Scores: The script provides individual toxicity scores for particular
toxicity categories, which goes a step further in providing a better knowledge of the evaluation for
those who want it. These ratings offer light on the type and level of toxicity that may be present in
the text. Users have access to a thorough assessment that provides them with insightful
knowledge to help them make wise decisions.

RESULTS:
In order to show the result of the model with the user input data and the model's accuracy in
classifying the user input as toxic or non-toxic language, I have used some offensive language
and civilized language, and the results are provided below:

Fig 5. : Result of user input data showing toxic language

13
 Fig 6. The result of the user input data shows non-toxic and toxic languages.

An essential component of our approach is the evaluation of user input, which allows for the
real-time screening of user-generated content for toxicity. It improves online conversation and
content control, adding to a safer and more civilized digital environment by offering transparent
results and optionally detailed scores. It's an important tool for users to make sure they have a
positive online experience.

14
 CHAPTER 4. CONCLUSION

In conclusion, this Python project for toxicity detection encompasses a strong and adaptable
instrument for identifying and reducing harmful language in user-generated material. This project
provides a comprehensive solution that includes real-time user input evaluation in addition to the
training of sophisticated toxicity detection models. It is appropriate for a variety of applications and
is a useful tool for maintaining a more courteous and safe online environment.
Users can choose to train a new model or use one that has already been taught, which allows for
flexibility and modification to meet different demands. For example, this model may be easily
included in post-API requests thanks to the API integration. By instantly screening content for
offensive or derogatory language, this integration can serve as a gatekeeper. Depending on the
findings, appropriate actions such as issuing warnings or content removal can be initiated, thereby
enhancing content moderation.
On the infrastructure side, it's imperative to have the model running as an API, ensuring rapid and
efficient responses to user requests on the server. This can be achieved through frameworks like
Flask or Django, making the model readily available on various online platforms. Whether it's for
social media platforms, forums, or any online space, the potential applications of this project are
vast and impactful.
The Model’s capacity to identify and regulate harmful language is a critical first step towards
promoting constructive interactions in a digital environment where upholding civilized and secure
online conversation is crucial. It is a prime example of how natural language processing, machine
learning, and web technologies are combined to make online communities safer and more
welcoming.

CHAPTER 5. FUTURE USAGE

This advanced toxicity detection algorithm has a wide range of intriguing applications, especially
when it comes to content filtering across many digital platforms. By automatically identifying and
removing harmful content, its inclusion into social media sites, forums, and online communities can
significantly contribute to the development of a more polite and safe online community. This not
only lessens the incidence of cyberbullying but also preserves valuable human resources for
moderation.
The adoption of this strategy will also assist educational institutions since it will allow teachers and
administrators to keep up a civil and supportive online dialogue on discussion boards and
classrooms. This instrument acts as a protector, warding off unwanted influences from positive
15
learning environments.
This technique can also be used by online publishers and news websites to control user-generated
material and discussions, which will encourage readers to have more purposeful and targeted
interactions. Apart from these uses, the model's flexibility is remarkable since it can be added to a
range of communication platforms, including email services and chat programs, providing an
automatic filter against offensive language.
Moreover, the need for efficient content filtering is expanding as the digital world changes, which
emphasizes how important this model's flexibility and agility are. Its capacity for multilingual
instruction creates opportunities for worldwide application, making it possible to establish an online
community devoid of hate on a grand scale. This strategy can create inclusive and polite digital
relationships, guaranteeing a good online experience for everyone.

16
 SECTION II. WRITTEN ASSIGNMENT
Data Visualization and Linear Regression Analysis in Python.

17
 1. INTRODUCTION

In data science and statistical modeling, linear regression analysis and data visualization are
fundamental methods. In this paper, we investigate these methods with a Python dataset that has
values for 'x' and 'y'. In order to understand the links between the variables, we will model the
relationship between 'x' and 'y1' using linear regression and show the results. Our goal in doing this
exploration is to show how effective these tools are for modeling and data analysis.

2. DATA ANALYSIS

2.1. Data Loading and Visualization

We start off by loading data from CSV files that are categorized into test, ideal, and training
datasets. This procedure makes use of the Pandas library. With the help of the Matplotlib software,
we create scatter plots to display the data. The links between the 'x' and 'y' numbers are made
clearer to us by these visuals. With 'x' as the common independent variable, we make distinct
scatter plots for 'y1', 'y2', 'y3', and 'y4'. These visualizations offer a basis in visual terms for
comprehending the properties of the data.

2.2. Linear Regression Analysis

In order to explore the data further and model a particular relationship, we utilize linear regression
analysis. Here, we are concentrating on 'x' and 'y1'. To initialize and fit a linear regression model to
the training set of data, we utilize the scikit-learn toolkit. We can learn more about the linear
relationship between 'x' and 'y1' by computing the model's coefficients, such as the intercept and
slope.

18
Fig 7. Data plot of x against y1, y2, y3 and y4 values

Fig 8. Scatter plot of x against y1.

19
 Fig 9. Scatter plot of x against y2.

Fig 10. Scatter plot of x against y3

20
 Fig 11. Scatter plot of x against y4.

2.3. Regression Metrics

It is essential to evaluate the model's effectiveness. We assess its accuracy using standard
regression metrics. Among these measurements are R-squared (R^2), Mean Absolute Error
(MAE), and Root Mean Square Error (RMSE). The prediction error of the model is quantified by
RMSE, the average prediction error is measured by MAE, and the percentage of variance
explained by the model is shown by R^2. These metrics offer a thorough evaluation of the linear
regression model's suitability for the training set of data.

Fig 12. Readings of RMSE, MAE AND R-Squared (R2) Values.

21
2.4. Visualizing Linear Regression

To provide a clear visualization of our linear regression model, we create a plot that showcases the
actual 'y1' values and the linear regression line. This visual representation helps us understand the
extent to which the model aligns with the training data. It serves as a valuable tool for
communicating the results of our analysis.

Fig 13. Plot of Linear Regression Model with actual value

22
 3. CONCLUSION
Data visualization and linear regression analysis are fundamental techniques in data science and
statistical modeling. This project has investigated their use in Python with a dataset that includes
values for 'x' and 'y'. We were able to model the relationship between 'x' and 'y1' through linear
regression, and we also obtained insights into the correlations between variables through data
visualization. We evaluated the model's performance using regression measures. This thorough
examination offers a template for applying these methods to other data analysis projects in addition
to offering insightful information.

Understanding data relationships, employing linear regression, and assessing model performance
are vital skills for data scientists and analysts. Data professionals can create predictive models and
make data-driven decisions with the help of the tools discussed in this article, which include
scikit-learn for linear regression analysis, Matplotlib for data visualization, and Pandas for data
manipulation. As this example shows, linear regression and data visualization are flexible and
effective methods for extracting meaning from data.

REFERENCE

1. Risch, Julian, Anke Stoll, Marc Ziegele, and Ralf Krestel. "hpiDEDIS at GermEval 2019:
Offensive Language Identification using a German BERT model." In Proceedings of the
GermEval 2019 Workshop, co-located with KONVENS 2019. Hasso Plattner Institute,
University of Potsdam, Heinrich Heine University Düsseldorf, University of Passau, 2019.

2. Thenmozhi, D., Senthil Kumar, B., Aravindan, Chandrabose, & Srinethe, S. (2019). NLP at
SemEval-2019 Task 6: Offensive Language Identification in Social Media using Traditional
and Deep Machine Learning Approaches. In Proceedings of the 13th International
Workshop on Semantic Evaluation (SemEval-2019) (pp. 739–744). Minneapolis,
Minnesota, USA. Association for Computational Linguistics.

3. Kaggle. (Year). Jigsaw Toxic Comment Classification Challenge Dataset. Retrieved from
https://www.kaggle.com/competitions/jigsaw-toxic-comment-classification-challenge/data

4. Sial, A. H., Rashdi, S. Y. S., & Khan, A. H. (2021). Comparative analysis of data
visualization libraries Matplotlib and Seaborn in Python. International Journal, 10(1).

23
 LIST OF APPENDICES

1. RESEARCH TOPIC CODE

2. WRITTEN ASSIGNMENT CODE

1. RESEARCH TOPIC CODE

24
25
26
2. WRITTEN ASSIGNMENT CODE

27
28
View publication stats

29