Spam Detection in Short Message Service

Using Natural Language Processing and

Machine Learning Techniques

MSc Research Project

Data Analytics

Anchal Ora
Student ID: x18135846

School of Computing
National College of Ireland

Supervisor: Dr. Catherine Mulwa

 National College of Ireland
Project Submission Sheet
School of Computing

Student Name: Anchal Ora

Student ID: x18135846
Programme: Data Analytics
Year: 2019
Module: MSc Research Project
Supervisor: Dr. Catherine Mulwa
Submission Due Date: 12/12/2019
Project Title: Spam Detection in Short Message Service Using Natural Lan-
guage Processing and Machine Learning Techniques
Word Count:
Page Count: 27

I hereby certify that the information contained in this (my submission) is information
pertaining to research I conducted for this project. All information other than my own
contribution will be fully referenced and listed in the relevant bibliography section at the
rear of the project.
ALL internet material must be referenced in the bibliography section. Students are
required to use the Referencing Standard specified in the report template. To use other
author’s written or electronic work is illegal (plagiarism) and may result in disciplinary
action.

Signature:

Date: 28th January 2020

PLEASE READ THE FOLLOWING INSTRUCTIONS AND CHECKLIST:

Attach a completed copy of this sheet to each project (including multiple copies). 
Attach a Moodle submission receipt of the online project submission, to 
each project (including multiple copies).
You must ensure that you retain a HARD COPY of the project, both for 
your own reference and in case a project is lost or mislaid. It is not sufficient to keep
a copy on computer.

Assignments that are submitted to the Programme Coordinator office must be placed
into the assignment box located outside the office.

Office Use Only

Signature:

Date:
Penalty Applied (if applicable):
 Spam Detection in Short Message Service Using
Natural Language Processing and Machine Learning
Techniques
Anchal Ora
x18135846

Abstract
As the usage of mobile phones increased, the use of Short Message Service
increased significantly. Due to the lower costs of text messages, people started
using it for promotional purposes and unethical activities. This resulted in the
ratio of spam messages increasing exponentially and thereby loss of personal and
financial data. To prevent data loss, it is crucial to detect spam messages as quick as
possible. Thus, the research aims to classify spam messages not only efficiently but
also with low latency. Different machine learning models like XGBoost, LightGBM,
Bernoulli Naı̈ve Bayes that are proven to be very fast with low time complexity
have been implemented in the research. The length of the messages was taken as
an additional feature, and the features were extracted using Unigram, Bigram and
TF-IDF matrix. Chi-Square feature selection was implemented to further reduce
the space complexity. The results showcased that Bernoulli Naı̈ve Bayes followed
by LightGBM with the TF-IDF matrix generated the highest accuracy of 96.5% in
0.157 seconds and 95.4% in 1.708 seconds respectively.
Keywords: Spam SMS, Text Classification, Natural Language Processing, Machine
Learning, Bernoulli Naı̈ve Bayes, LightGBM, XGBoost

1 Introduction
Short Messaging Service (SMS) is mainly used for unofficial communication such as pro-
moting new products and services but at times also used for official communication like
information about any bank transaction or confirmation of the order on an online portal
etc. Due to advancements in technology, the costs of sending an SMS have reduced
drastically. This has proved to be a boon for some whereas a bane for many. People
are misusing the SMS facility to promote products, services, offers and schemes and so
on. How annoying this has become can be assessed by the fact that people have started
ignoring SMS they receive because twenty to thirty percent of the total SMS received is
spam (Kim et al. (2015)). This menace is growing at a rapid rate. As a result, people
miss out on genuine informative messages such as bank transactions. At times the ig-
norance towards SMS can prove detrimental because some fraud transactions might have
been performed, but the information was neglected. The motive behind this project is to
apply machine learning algorithms to separate spam messages from genuine ones. Ma-
chine learning techniques along with Natural Language Processing techniques was used
to make the process more agile and efficient.

1
1.1 Motivation and Background
For today’s generation, usage of phones is not just confined to communication now but an
array of different uses such as storing personal information like documents, notes, media,
making financial transactions, shopping, etc. Owing to a wide range of information stored
on devices some of which are personal and critical, hacking phones is of utmost interest to
people having unethical intentions. SMS is an easy way to target people because it is used
by people from all walks of life and all ages and they are not aware of the implications.
Hackers access phone and all its information when a phone is hacked, and people have
absolutely no idea about it. Consequently, there is a loss of critical data that can be
exploited for illegal purposes. It can be traumatic for the victims causing psychological
despair and financial loses.
Messages are not just written in English but in various languages and those in English
can also have words and abbreviations which are from other languages. Hence identifying
spam SMS is a challenging work as stated by (Yadav et al. (2012)) since it does not have
any header like in case of emails. So, the techniques used to figure out spam emails cannot
be used for messages. New solutions must be devised for new problems and thus many
researchers have been working in this area to develop new algorithms and techniques. In a
recently concluded research by (Gupta et al. (2018)), a deep learning algorithm has been
implemented with distinct traditional algorithms. Through this, the researcher was able
to achieve an accuracy of 99.1 percent which is the best result achieved so far. According
to (Yue and Elfayoumy (2007)), implementation of deep learning models comes with
challenges. It is an arduous task to apply deep learning models to actual datasets since
it becomes computationally expensive and time-consuming. Although a model generates
accurate results, it takes time which is not good enough in case of SMS because people
tend to access the SMS within seconds of receiving it. This gap in research demands
the creation of a model that performs efficiently with high accuracy consuming very less
time. Hence, below research question is put forward in this research.

1.2 Research Question

Spam SMS becomes very irritating and disappointing to the mobile phone users. This
can lead to crucial information loss by clicking on any spurious links or even the genuine
information could be missed by ignoring the actual message as spam. Hence there is a
need to design a model that is not only efficient but quick in the detection of spam SMS.
So the Research Question is:
RQ: “To what extent can classification of spam SMS using a combination of Nat-
ural Language Processing techniques (Bag of Words and TF-IDF) and machine learning
techniques (XGBoost, LightGBM, SVM, Bernoulli Naı̈ve Bayes, and Random Forest)
improve the efficiency in detection of spam SMS with low latency to help mobile phone
users?”
This research helps in filtering the spam SMS quickly and efficiently which can prevent
data loss. In this research, the problem area is tackled by implementing models like
Bernoulli Naı̈ve Bayes, SVM, Random Forest and boosting techniques like XGBoost and
LightGBM. Gradient Boosting technique has proven to be very efficient, flexible and
quick in generating results from actual datasets and has won accolades on several data
mining platforms (Tianqi and Guestrin (2016)). Along with these techniques, feature
selection technique Chi-Square is implemented which further reduces dimensionality and

2
in turn decreases time complexity. To make the model more reliable, Stratified 10-Fold
Cross-Validation technique was used.
To solve the research question, the research objectives are implemented as depicted
in section 1.3 (Table 1).

1.3 Research Objectives and Contributions

One of the important objectives was to do an investigation on the work already done in
the field of text analytics which helped in gaining the knowledge and achieving the below
research objectives.

Table 1: Research Objectives

Objectives Description Evaluation Metrics

Obj-1 Spam SMS file dataset pre-processing

Obj-2 Implement Natural Language Pro-
cessing techniques
Obj-2a Implement Bag-Of-Words to create doc-
ument term matrix for Spam SMS De-
tection
Obj-2b Implement TF-IDF to create document
term matrix for Spam SMS Detection
Obj-3 Apply feature selection technique i.e.
Chi-Square method
Obj-4 Implement, Evaluate and generate the
outcome for several machine learning al-
gorithms
Obj-4a Implementation, Evaluation, and Res- Accuracy, Precision, Recall,
ults of XGBoost F1-Score, Execution Time
Obj-4b Implementation, Evaluation, and Res- Accuracy, Precision, Recall,
ults of LightGBM F1-Score, Execution Time
Obj-4c Implementation, Evaluation, and Res- Accuracy, Precision, Recall,
ults of Bernoulli Naive Bayes F1-Score, Execution Time
Obj-4d Implementation, Evaluation, and Res- Accuracy, Precision, Recall,
ults of SVM F1-Score, Execution Time
Obj-4e Implementation, Evaluation, and Res- Accuracy, Precision, Recall,
ults of Random Forest F1-Score, Execution Time
Obj-5 Compare the developed models(Obj-5a Accuracy, Execution Time
to Obj-5e)
Obj-6 Compare the developed models with the Accuracy
existing state-of-the-art models

3
 Table 2: Research Objectives
Objectives Description Evaluation Metrics

Obj-1 Spam SMS file dataset pre-processing

Obj-2 Implement Natural Language Pro-
cessing techniques
Obj-2a Implement Bag-Of-Words to create doc-
ument term matrix for Spam SMS De-
tection
Obj-2b Implement TF-IDF to create document
term matrix for Spam SMS Detection
Obj-3 Apply feature selection technique i.e.
Chi-Square method
Obj-4 Implement, Evaluate and generate the
outcome for several machine learning al-
gorithms
Obj-4a Implementation, Evaluation, and Res- Accuracy, Precision, Recall,
ults of XGBoost F1-Score, Execution Time
Obj-4b Implementation, Evaluation, and Res- Accuracy, Precision, Recall,
ults of LightGBM F1-Score, Execution Time
Obj-4c Implementation, Evaluation, and Res- Accuracy, Precision, Recall,
ults of Bernoulli Naive Bayes F1-Score, Execution Time
Obj-4d Implementation, Evaluation, and Res- Accuracy, Precision, Recall,
ults of SVM F1-Score, Execution Time
Obj-4e Implementation, Evaluation, and Res- Accuracy, Precision, Recall,
ults of Random Forest F1-Score, Execution Time
Obj-5 Compare the developed models(Obj-5a Accuracy, Execution Time
to Obj-5e)
Obj-6 Compare the developed models with the Accuracy
existing state-of-the-art models

Contributions: The major contribution resulting from this research was the imple-
mentation of machine learning models such as XGBoost, LightGBM, Random Forest,
SVM, and Bernoulli Naı̈ve Bayes that are capable to detect spam SMS with good effi-
ciency and less time.
The minor contribution of the research was the comparison of results of fully developed
models along with the existing models on evaluation metrics using the visualizations.
These contributions will help mobile phone users as it will effectively resolve the problem
of spam SMS by detecting it with the minimum time.

The rest of the technical report contains the following chapters: Chapter 2 represents Lit-
erature Review, Chapter 3 consists of Scientific Methodology and Design Specifications,
Chapter 4 talks about Data Pre-processing, Implementation, Evaluation, and Results of
Spam Short Message Service Classification Models, Chapter 5 contains Discussion and
Comparison of Results and lastly, Chapter 6 consists of Conclusion and Future Work.

4
2 Literature Review
2.1 Introduction
SMS Spam is a serious problem and thus many researchers have keen motivation to
solve the problem using different approaches and methods. This section investigates the
problem of spam and various other text mining problem and how it is tackled by imple-
menting different machine learning algorithms and techniques by reviewing the literature
from the year 2007 to date. The review is divided into the following subsections: (i)A
Critique of Spam Short Message Service Detection (ii)A Critical Review of Algorithms
and Techniques Used in Spam Classification and Identified Gaps (iii) A Critique on Data-
set and Features used in Short Message Service Spam Detection (iv)A Comparison of the
Reviewed techniques.

2.2 A Critique of Spam Short Message Service Detection

Over the past few years, mobile phones have encountered an enormous amount of growth
and therefore SMS has become the common platform for the exchange of information. It
was reported that there are around 3.5 billion active mobile phone users in the year 2010
(Shirani-Mehr (2012)). The researchers also mention that SMS had gained popularity
among the youth and as the SMS charges reduced over the years, around 30% of total
messages were found to be spam in Asia in the year 2012. The number of spam SMS
detection software available is very limited, which makes spam SMS detection an inter-
esting problem that can be solved by using machine learning techniques.
As per the researcher (Fernandes et al. (2015)), the short message service was started in
the year 1992 and has gained substantial popularity with revenue close to 153 billion dol-
lars in the year 2016. A humongous amount of SMSes are exchanged all over the world on
a daily basis. Several misuses of SMSes have been highlighted leading to loss of personal
and financial information by unknowingly clicking on bogus links. The researcher (Sethi
et al. (2018)) revealed a very appalling fact that the amount of spam SMS circulated
the world over surpass the number of spam emails. The problem is so serious that some
countries like Japan took legal action against it. One of the main reasons for spam SMS
is because the cost of sending an SMS is so low that it becomes irrelevant compared to
the benefit hackers can reap if they get hold of the sensitive information.
Researcher (Agarwal et al. (2016)) mentions that spam is the junk or unwanted messages
which are broadcasted to a group of users with bad intention like stealing personal in-
formation, etc. The spam messages are growing at a rapid rate with almost 500% increase
every subsequent year. The researcher also mentions that the quantum of spam SMS is
not the same across the regions. For instance, in North America only one percent of total
SMS were spam in the year 2010, however, more than thirty percent were spam in some
parts of Asia. In the year 2008, the spam messages received by the people in China in a
week were around 200 billion.
Looking at the above-stated facts mentioned in research papers regarding the growth of
spam SMS, it is high time to solve the problem using the latest techniques which are
efficient and can be implemented in real-time.

5
2.3 A Critical Review of Algorithms, Techniques Used in Spam
Classification and Identified Gaps
2.3.1 A Review of Short Message Service Spam Classification
An analysis of spam SMS filtering was done on the UCI machine learning dataset in the
year 2015 by (Kim et al. (2015)) who chose the frequency ratio feature selection tech-
nique while implementing the algorithms like Naı̈ve Bayes, Logistic Regression and J-48
Decision Trees where the 10 fold cross-validation technique was applied. It was seen that
Naı̈ve Bayes generated results in a minimum time with the highest accuracy of 94 per-
cent. In the year 2018, a similar analysis was conducted by (Gupta et al. (2018)) using
2 sets of data, one with UCI machine learning which is the same corpus as of Kaggle
with a total of 5574 ham and spam messages and another dataset contains 2000 spam
and ham messages. TF-IDF matrix was created and then machine learning algorithms
like Naı̈ve Bayes, Random Forest, SVM, Decision Tree, Convolutional Neural Network
and Artificial Neural Network were applied on both the datasets. The results obtained
by CNN were the state-of-art in this area with an accuracy of 99.10 followed by Naı̈ve
Bayes and SVM.
A research conducted by (Ma et al. (2016)), on spam SMS detection proposes a message
topic model which is a form of probability topic model. It uses the KNN algorithm to
remove the sparsity problem in the messages. The symbol terms, background terms were
considered and it was found that the model generated better results than the standard
LDA. The classifier GentleBoost was used for the first time in the research done by (Ak-
bari and Sajedi (2015)) on SMS spam detection in the year 2015. For unbalanced data
and binary classification, boosting algorithms work well. GentleBoost is a combination
of two algorithms, namely AdaBoost and LogitBoost. GentleBoost is well known for its
higher accuracy and less consumption of storage as it removes unwanted features. It
obtained an accuracy of 98% on the dataset consisting of 5572 text messages.
The author (Agarwal et al. (2016)) states that the short length of the messages and the
use of casual words in the text messages do not allow it to perform well with the already
established solutions of email spam filtering. In this research, it can be seen that SVM
followed by Multinomial Naı̈ve Bayes (MNB) shows outstanding results in terms of ac-
curacy with 98.23 and 97.87% respectively. MNB took the least execution time of 2.03
seconds. The researcher further suggests that features like the number of characters in
the messages or definite threshold to the length of the message can increase the perform-
ance.
Looking at the above research, it can be stated that the performance of traditional al-
gorithms like Naı̈ve Bayes and SVM is superior to other algorithms. Also in this research,
the length parameter can be taken as an additional feature to check if it enhances the
performance of the model.

2.3.2 A Review of Email Spam Classification

In the text classification area, the Naı̈ve Bayes algorithm is very popular in generating
good results. The research conducted by (Almeida et al. (2011)), on email spam filtering
used 7 different versions of Naı̈ve Bayes such as Basic Naı̈ve Bayes, Bernoulli Naı̈ve Bayes,
Gaussian Naı̈ve Bayes to name a few. The feature selection techniques applied were Chi-
Square and Mutual Information. The model implementing Bernoulli Naı̈ve Bayes in
combination with Chi-Square generated better results.

6
Significant power and memory are needed if the number of features in a classification
problem is excessive. Higher the features, more is the dimensionality and greater is the
need for power and memory. By removing the features that are not pertinent or redundant
and pulling out the beneficial ones, performance can be amplified as stated by the paper
(Ergin and Isik (2014)) which is on email spam filtering. Eliminating the stop words,
stemming and normalizing the dataset is followed by the creation of Document Term
Matrix using Bag of Words. Techniques like Gini Index, Chi-Square and Information
Gain were used for feature selection on machine learning algorithms like Artificial Neural
Networks (ANN) and Decision Tree. An outstanding result was generated by the coalition
of Bag of Words and Chi-Square techniques on ANN with the accuracy as 91%.
In the research performed by (Islam et al. (2009)) on the spam filtering techniques
such as Naı̈ve Bayes and Artificial Neural Networks (ANN) were applied. The features
from the header and the body of an email were taken into consideration. It was found that
Naı̈ve Bayes outperformed ANN with higher accuracy, recall, and precision. A paper by
(Lee et al. (2017)) found that Weighted Naı̈ve Bayes is not only computationally effective
but also very efficient in case of spam detection even with new destructive campaigns. The
evaluation was carried on eight datasets that were sourced from two sites. The accuracy
attained is around 95% for both the sources. The researcher (Yue and Elfayoumy (2007))
argues that despite neural networks producing good results, it takes a huge amount of
time in generating results. Also, once the model is built it stops learning from new
emails, unlike Naı̈ve Bayes which is adaptive and trains from new emails. The researcher
suggested that applying techniques like boosting in the future which can produce more
quick results.
The researcher (Yu and ben Xu (2008)) discusses the pitfalls of email spams as these
messages not only waste time and energy of the end-users but lead to issues like utilizing
high mailbox space and bandwidth. The researcher tried to solve this problem by imple-
menting 4 machine learning models like Support Vector Machine, Naı̈ve Bayes, Neural
Network and Relevance Vector Machine. The performance of the models was computed
using the training set of varying sizes and the results demonstrated that Neural Networks
are not suitable for spam filtering as they are susceptible to the size of training data.
SVM and RVM performed better where SVM took less training time than RVM.
Looking at this section, it can be interpreted that Naı̈ve Bayes along with feature
selection like the Chi-Square technique shows great results in text classification. The
results produced are not only better in terms of accuracy but also in terms of time
complexity. Also, the deep learning techniques like Neural Networks were not found to
be productive in spam detection as it is sensitive to the size of the training data. It takes
a large amount of time to train and test the dataset which makes them computationally
ineffective.

2.3.3 An Investigation of Other Text Classification Areas

Phishing is a type of cybercrime where sensitive information such as user credentials, card
details, etc. are stolen through counterfeit websites or emails. The paper by the researcher
(Li et al. (2019)) provided the solution by implementing Word2Vec as a feature extraction
technique that gathers the feature from HTML code. Along with the feature extraction, a
joint model using algorithms like LightGBM, XGBoost, and Gradient Boosting Trees are
built which raised the performance to 97% accuracy. The researcher (Liew et al. (2019)),
detected phishing tweets on a real-time basis utilizing the feature from the phishing

7
URL. Random Forest technique was implemented and found to produce an accuracy of
94.5%. The researcher (Koray et al. (2019)) detects the phishing websites by analyzing
and extracting features from the URL. The Random Word Detection module is used
which decomposes URL into small features which are then used to classify if the websites
are legitimate or not. Seven different machine learning algorithms like Naı̈ve Bayes,
KNN, SVM, Random Forest, Decision Tree, Adaboost, K-star were implemented on a
humongous amount of data. It was seen that Random Forest produced the highest
accuracy of around 97.98 percent among all the techniques applied.
In the year 2018, the researcher (Yuan et al. (2018)) proposed a blend of features
pertaining to URL and web page for the detection of phishing websites. Along with the
basic features like the length of the URL, unusual characters, etc, statistical features like
mean, median and variance and lexical features like title and the content of the web page
were also considered. Several algorithms like KNN, Logistic Regression, Random Forest,
Deep Forest, XGBoost were applied. It was found that Deep Forest followed by XGBoost
manifested high accuracy and less training time.
Nowadays, the success of any business heavily depends on authentic reviews. How-
ever, not all reviews are authentic. The ratio of genuine reviews to sham reviews varies
significantly from case to case and higher the bogus reviews, more is the image maligned
of a business. Implementing sentiment analysis using natural language processing tech-
niques can effectively detect the opinion spams. Earlier, researchers like (Jindal and
Liu (2007)), (Ren and Ji (2017)) used supervised and semi-supervised algorithms for the
detection of spam opinions. These models have a few restrictions such as low flexibil-
ity, high computational time and poor accuracy. These limitations were overcome by
the researcher ((Hazim et al. (2018)) using the Gradient Boosting models like XGBoost,
GBM Bernoulli, GBM Adaboost and GBM Gaussian. The opinion spam detection was
performed on multilingual datasets. It was found that XGBoost outperformed the other
models for the English language dataset generating high recall percentage whereas GBM
Gaussian produced good results for the Malay language dataset. The researcher (Prieto
et al. (2016)) detects opinion spam using neural networks and it was observed that model
complexity increases due to the large set of details provided to the neural network and
thus increases the overall computational cost.
In the world of online advertising, fraud clicks are one of the most momentous issues.
The research done by (Minastireanu and Mesnita (2019)) tackles the problem of fraud
clicks by using the latest machine learning technique viz. LightGBM on the dataset which
contains millions of clicks. The K-Fold Cross-Validation technique is used as a feature
engineering which helps in improving the performance. The accuracy achieved by the
model was 98% and was found to be the fastest with respect to computational speed
and low on memory consumption. Looking at the above research papers, it was found
that LightGBM and XGBoost are suitable as it performs faster with a less computational
speed, unlike the deep learning techniques. Also, Random Forest performed well giving
high accuracy and hence in this research, these algorithms are chosen.

2.4 A Review of Datasets and Features used in Short Message

Service Spam Detection
Spam SMS is a serious problem in Vietnam due to the cheap message pre-paid bundle
available. The research done by (Pham (2016)) focuses on the detection of spam SMS for
the Vietnamese language. The dataset contains 6599 messages marked as ham or spam.

8
Out of these 1042 were spam and rest are legitimate. The length feature was added
as the spam messages tend to be longer than ham. The data was pre-processed using
NLP techniques and then Bag of Words (BoW) and Term Frequency - Inverse Document
Frequency (TF-IDF) were chosen as feature selection techniques. A content-based spam
message filtering method proposed by (Balli and Karasoy (2018)) having a dataset of 5574
messages of which 747 are spam uses the semantic relationship between the SMS words.
The feature selection is done using the Word2Vec algorithm which calculates the distance
between the vector of the words and thus features are extracted for each message. The
work presented by the researcher (Aich et al. (2019)) for spam detection on imbalanced
datasets of SMS, implemented the SMOTE approach which generated great results in
combination with the SVM algorithm and the accuracy increased by seven points.
The researcher (Najadat et al. (2014))investigates spam SMS detection by implement-
ing 12 different types of classifiers on the dataset which contained 5574 messages. The
research used the technique of down-sampling where the ham messages were reduced
to the count of spam messages. The main contribution of this research is to examine
the impact of class imbalance on performance and the researcher found that a balanced
dataset produces more accurate results. The results are compared to the previous re-
search done with the imbalanced dataset and found that performance degrades due to
the under-fitting issue.
It can be seen that the datasets which have the issue of class imbalance can lead to bias
results and poor performance whereas the balanced dataset can improve the performance.
Also, feature selection techniques like Bag Of Words and TF-IDF works well in the case
of text classification.

2.5 A Comparison of the Reviewed techniques

A high-level comparison of the related work in text analytics classification problem based
on feature used and classifiers applied. It can be seen that classifiers like Naı̈ve Bayes,
Random Forest, LightGBM, XGBoost, and SVM outperform other classifiers and hence
these are chosen to be implemented in the research.

9
 Table 3: Comparison of Reviewed Features and Classification Techniques
Area of Applied Classi- Features Ex- Results Author
Classifica- fiers tracted
tion
Spam SMS Naive Bayes, J- Message-Keyword Naive Bayes (Kim et al.
Classification 48, Logistic Re- Matrix (Bag of with 94.7% (2015))
gression Words) accuracy
Spam SMS Naive Bayes, TFIDFVectorizer CNN with (Gupta et al.
Classification SVM, Random 99.1% accur- (2018))
Forests, Ada- acy
boost, CNN,
ANN,Logistic Re-
gression, Decision
Tree
Email Spam ANN, Na¨ıve Multiple features Naive Bayes (Islam et al.
Classification Bayes from header and with 92% (2009))
body like URL, with low ex-
images, links, etc ecution time
than Neural
Networks
Phishing Logistic Regres- TF-IDF for fea- Deep Forest (Yuan et al.
Classification sion, Decision ture extracted with 97.7% (2018))
Tree, GBDT, from URL and and XGBoost
Deep Forest, links with 97.1%
Random Forest,
XGBoost, KNN
Phishing Naı̈ve Bayes, NLP features, Random (Koray et al.
Classification KNN, SVM, word vectors Forest with (2019))
Random Forest, 97.98%
Decision Tree,
Adaboost, K-star
Fraud Advert- LightGBM, XG- IP, OS, channel, LightGBM (Minastireanu
isement Clas- Boost. Stochastic device, click time with 98% and Mesnita
sification Gradient Boosting accuracy (2019))

2.6 Conclusion
Looking at the related work done in the area of text classification, it is clearly seen
that gradient boosting techniques produce quick and accurate results. As suggested, the
length feature can be a useful parameter in deciding the type of SMS. Hence the same
was implemented in this research. There is an urgent need for developing a spam SMS
detection model by combining the best-reviewed machine learning techniques with the
NLP techniques which can generate good results and can answer the research question
(section 1.2) and the research objectives (section 1.3). The next chapter discusses the
scientific methodology and design specifications chosen to develop the spam detection
model which helps mobile phone users.

10
3 Scientific Methodology and Design Specifications
In data mining projects, different methodologies are used. The most common ones are
CRISP-DM, KDD, and SEMMA. Knowledge Discovery and Data Mining (KDD) suits
well for the research as the deployment step is not required. KDD is a very precise and
complete approach that focuses not only on the business process but also on implement-
ation(Shafique and Qaiser (2014)). The design architecture is a two-tier which contains
a Presentation Layer and Application Layer.

3.1 Methodology for Short Message Service Spam Detection

The methodology for Short Message Service Spam Detection (Figure 1) is an iterative
process that has the following phases :
(a) Data Selection phase where the UCI machine learning repository provided data to
Kaggle, and the data was downloaded from Kaggle which is in .csv format.
(b) Data cleaning and preprocessing was done in python using the NLTK library.
(c) The data transformation was done by normalization and creation of Document Term
Matrix followed by feature selection using a statistical technique like Chi-Square.
(d) In the data mining step, machine learning models such as XGBoost, LightGBM,
Bernoulli Naı̈ve Bayes, SVM, and Random Forest were implemented.
(e) All the models were evaluated and interpreted on the basis of different parameters
like Accuracy, Precision, Recall, F1-Score and Execution Time.

Figure 1: Methodology for Short Message Service Spam Detection

3.2 Project Design Specification

As depicted in Figure 2, the design process for the Short Message Service Spam Detection
model consists of two-tier that is the Presentation Layer and the Business Logic Layer.
The reason for choosing 2-tier architecture over 3-tier architecture was the format of the
dataset. The data was already in a structured format and did not require any additional
conversion apart from the pre-processing. The presentation layer consists of interpretation
of results in the form of visualizations which was done in the Microsoft Power BI tool. In
the Business Logic Layer, data was fetched from Kaggle which was then pre-processed,
transformed, trained and evaluated using different classification models.

11
 Figure 2: Design Process for Short Message Service Spam Detection

Google cloud platform Colab was used for the implementation of the project.
Colab is a free to use platform as a service provided by Google. It is the preferred
platform of a large number of developers for machine learning since it has almost all the
packages pre-installed and does not require any special installations.
The implementation, evaluation, and results are discussed in detail in the next chapter.

4 Data Pre-processing, Implementation, Evaluation

and Results of Spam Short Message Service Clas-
sification Models
4.1 Introduction
The section discusses all the steps involved in the implementation along with evaluation
and results. Figure 3 illustrates the Workflow Diagram for the Short Message Service
Spam Detection.The Workflow of the project is explained in detail step by step in this
section.

12
 Figure 3: Work Flow Diagram for Short Message Service Spam Detection

4.2 Exploratory Data Analysis

Exploratory Data Analysis is a crucial process where the data is analyzed to uncover
underlying patterns, spot abnormality and test the hypothesis . It is the best practice
to understand the data and then carry out the data mining process. Missing value
analysis was carried out using libraries of python like Pandas. An additional feature of
the “length of text messages” was considered as there was a substantial correlation of
over 0.6 between the length and the type of SMS as seen in Figure 4. For visualization
of most frequent words appearing in spam messages and ham messages, the Matplotlib
library with WordCloud technique was used (Refer Figure 5 and 6).

13
 Figure 4: Correlation Matrix For
SMS Type and Length Feature

Figure 5: Word Cloud for Spam Figure 6: Word Cloud for Ham

4.3 Data Sampling

Class imbalance is most commonly observed in machine learning classification problems.
It occurs where the 2 classes have disproportionate data. Under such circumstances, most
of the algorithms generate biased results more towards the majority class.( Longadge and
Dongre (2013)) The problem can be solved by implementing 2 techniques that are either
up-sampling or down-sampling. Up-sampling is a technique where the minority class
data is artificially created or duplicated and reached to the level of the majority class.
Up-sampling is majorly done when the minority class is highly low in number.The other
technique that is down-sampling is where the majority class samples are reduced to the
same level as minority class sample size. In this research down-sampling is done using
random sampling as minority class samples were in decent number. To implement this
technique no external libraries were imported. Figure 7 shows the data before and after
down sampling.

14
 Figure 7: Message Count before and after Down Sampling

4.4 Data Pre-processing

Data pre-processing is mainly cleaning of data by removing unwanted rows, columns,
missing values, outliers, etc. For the research, the following pre-processing steps have
been taken:
Removal of Unwanted Columns It was found that there were 3 extra columns without
data in it, which adds extra noise to the model hence removed.
Cleaning of Text Messages For cleaning the text messages, the following steps are
involved:
•Tokenize the words: In this step, the words are split into tokens based on white spaces
or punctuation. To achieve it, word tokenize function from NLTK library was used.
•Removal of stop words from text messages: Stop words are basically the most
commonly used words (such as “a”, “an”, ”the”) which increases the dimensionality and
impacts the efficiency of the model (Vani and Gupta (2014)). Stopwords package from
NLTK library was used where the stop words were removed if they appear in the pre-
defined stop words list, thereby filtering the text messages.
•Lemmatizing words: Lemmatization in simple terms refers to the removal of duplic-
ate data. For example, words like “study”, “studying”, and “studies” are considered 3
different words after the creation of a Document Term Matrix and hence increases the
dimensionality. WordNetLemmatizer package from the NLTK library was implemented
which helped in converting the inflected word to its base format.
•Normalize the words: The word-stock can be decreased if all the words are in either
lowercase or uppercase. For example, “OFFER” and “offer” are considered 2 separate
words before normalization. The lower() function in python helped in achieving the nor-
malization of data.
Encoding the categorical column LabelEncoder() function from the skelearn.preprocessing
library is used to encode the categorical dependent column.The label encoding is preferred
over one-hot encoding because the SMS type is a 2-class column and not multi-class, which
is either spam or ham.

15
4.5 Document Term Matrix using NLP Techniques
As the machine learning models work only on mathematical data, a matrix was created
which contains the word and its frequency of occurrence. The following two techniques
were used for the creation of the document term matrix.
Bag of Words (BoW): Bag of Words is a way of extracting the features from the set of
text messages. In this, a matrix called a bag of words is created (shown in Figure 8) which
describes the text based on the frequency of the word appearing in the document.This
was implemented using the CountVectorizer package in python. The CountVectorizer can
be implemented in the form of n-grams. In the research, the feature was extracted using
Unigram and Bigram matrix. The Unigram matrix comprises a single word whereas the
Bigram matrix consists of two consecutive words from a document.

Figure 8: Bag Of Words Matrix

Term Frequency-Inverse Document Frequency (TF-IDF) : Term Frequency-

Inverse Document Frequency commonly known as TF-IDF is one of the common tech-
niques used to represent the text data into vector form. This matrix (shown in Figure
9) helps in understanding the importance of the word in the corpus of documents. Fre-
quently occurring term in a given document that seldom occurs in other documents in
the corpus has high TF-IDF value. To implement the TF-IDFVectorizer() function was
used with the n-gram as a unigram. The machine learning models were implemented
after the creation of the matrix.

Figure 9: TF-IDF Matrix

4.6 Feature Selection

It is one of the most important steps in the area of text mining. It is the process of
selecting the important attributes which contribute the most in making the classifica-
tion. Feature Selection can enhance the training time of a model. It reduces over-fitting
and thus increases the accuracy and performance of the model. The more the features,
the more complex a model, hence it reduces the complexity and is easy to understand.
There are many techniques used for feature selection. For the research, the Chi-Square
Technique was implemented.
Chi-Square Feature Selection Technique: Chi-Square is a statistical test which elim-
inates the less important features by checking the relationship between dependent and

16
independent feature. It checks the deviation of the observed count from the expected
count. Chi-Square technique performs well in the case of categorical data and the Bag-
Of-Words and TF-IDF matrix consist of categorical data only and therefore the technique
is preferred in the research. Also, from the literature (Ergin and Isik (2014)) and (Al-
meida et al. (2011)) it can be seen that Chi-Square generated excellent outputs in the
field of text analytics. Chi2 package from the sklearn library was used for selecting the
best features.
This accomplished the research objective from Obj-1 to Obj-3 mentioned in Chapter 1,
Section 1.3.

4.7 Implementation, Evaluation, and Results of Spam Short

Message Service Classification Models
After the selection of features, machine learning models like Naı̈ve Bayes, Random Forest,
XGBoost, LightGBM, and SVM. The evaluation of the models was done using a Strat-
ified 10-Fold cross-validation approach on the basis of Accuracy, Precision, Recall,
F1-Score and Execution Time. The accuracy and execution time being the most signific-
ant and were used to solve the research question. As mentioned by (Gupta et al. (2018))
choosing the correct metric for evaluation is important in interpreting the results from
the observations. The following parameters which help in better understanding of the
evaluation metrics which is considered in the research:
a. True Positive(TP): The positive class is correctly predicted by the model
b. True Negative (TN): The negative class is correctly predicted by the model
c. False Positive (FP): The model predicts the class as positive but it should be neg-
ative
d. False Negative (FN): The model predicts the class as negative but it should be
positive

Accuracy: It measures how close the observed value is from the actual value. As
per (Aich et al. (2019)) the classification accuracy metric is more transparent when the
classes are balanced. It is formulated as:

Accuracy = (TP+TN)/(TP+TN+FP+FN)

Precision: It is the ratio of the correctly predicted positive results to the total num-
ber of positive results that a model predicts. It can also be interpreted as how much the
model is relevant when it predicts. (Gupta et al. (2018)) It is formulated as :

Precision = (TP)/(TP+FP)

Recall: It is the true positive rate or sensitivity. It is the proportion of correctly

predicted positive instances to the total number of instances in the actual class. Recall
shows the potential of a model to find all the positive instances. It can be formulated as:

Recall = (TP)/(TP+FN)

F1-Score: F1-Score is the weighted average of recall and precision. It measures the
incorrectly classified instances. It is formulated as:

17
 F1-Score = 2*((Precision*Recall)/(Precision+Recall))

Stratified 10-Fold Cross-validation Technique: It is a statistical method where

the data was divided into 10 folds in such a way that each fold represents whole data.The
data was reshuffled after every fold and was repeated 10 times for this research. This
technique helped in reducing the variance as different training and testing sets were used.
After the cross-validation, the average of all the evaluation parameters were taken.

4.8 Experiment 1: Implementation, Evaluation and Results of

XGBoost Model
4.8.1 Implementation
Extreme Gradient Boosting is an ensemble boosting machine learning technique based on
the decision-tree algorithm and works in sequential manner. As per the researcher,(Tianqi
and Guestrin (2016)) the model generates output very quickly compared to traditional
machine learning algorithms. The other advantages of XGBoost are:
It has the capability to handle sparse data like BoW matrix.It uses parallelization which
helps it to generate the results faster and suitable for low latency applications.
XGBoost Classifier was implemented using Stratified 10-Fold Cross-Validation to di-
vide the dataset into training and test sets. The function used to implement the model
was the XGBClassifier() of the sklearn library. The model was implemented on Unigram,
Bigram, and TF-IDF with and without length feature.

4.8.2 Evaluation and Results

Looking at the Figure 10, it was evident that for the XGBoost model, there was a
performance upgrade on taking the length feature into consideration. An increase in
Accuracy, Recall, and F1-Score was depicted by the length feature. Length feature had
a high impact on the Bigram matrix. The model performed well with the Unigram and
TF-IDF matrix giving an Accuracy of 0.944, F1-Score of 0.943, and Recall of 0.929.
Precision for Bigram without length feature was the highest at 0.997.

Figure 10: Evaluation Metrics for XGBoost Model

18
4.9 Experiment 2: Implementation, Evaluation and Results of
LightGBM Model
4.9.1 Implementation
Light Gradient Boosting machine is recently developed machine learning model in 2017.
It is a boosting technique and similar to XGBoost with advantages like high efficiency,
low memory usage, fast training speed, better accuracy, capable to handle large datasets
and support parallelization 1 . As it is new, comparatively less research has been done on
it but looking at the literature, it can be seen that it performed well in the field of text
analytics. (Minastireanu and Mesnita (2019) (Li et al. (2019))
LightGBM classifier was implemented using Stratified 10-Fold Cross-Validation to di-
vide the dataset into training and test sets. The function used to implement the model
was the LGBMClassifier() of the sklearn library. The model was implemented on Uni-
gram, Bigram, and TF-IDF with and without length feature.

4.9.2 Evaluation and Results

As observed in the Figure 11,the LightGBM model generated nearly the same accuracy for
the TF-IDF and Unigram matrix. The length feature contributed the most to the Bigram
matrix where the performance was enhanced and there was a significant improvement in
the Accuracy, Recall, and F1-Score. It can be seen that the model performed the best
with the TF-IDF matrix generating an Accuracy of 0.957, Precision of 0.958, Recall of
0.956 and F1-Score of 0.957.

Figure 11: Evaluation Metrics for LightGBM Model

4.10 Experiment 3: Implementation, Evaluation and Results of

Bernoulli Naı̈ve Bayes Model
4.10.1 Implementation
Naı̈ve Bayes is a simple algorithm which works on conditional probability and is mostly
used in classification problems. The results generated by the researcher (Almeida et al.
(2011)), (Lee et al. (2017)) showed that Naı̈ve Bayes not only performed well in terms
of accuracy and time.Bernoulli Naı̈ve Bayes is a different form of classical Naı̈ve Bayes
1
https://lightgbm.readthedocs.io/en/latest/

19
and it works well with categorical data(Xu (2018)). Therefore it makes the classifier a
suitable choice to implement in the research.
Bernoulli Naive Bayes classifier was implemented using Stratified 10-Fold Cross-Validation
to divide the dataset into training and test sets. The function used to implement the
model was BernoulliNB() of the sklearn library 2 . The model was implemented on Uni-
gram, Bigram, and TF-IDF with and without length feature.

4.10.2 Evaluation and Results

From the Figure 12, it can be seen that the Bernoulli Naı̈ve Bayes model generated the
highest accuracy of 0.965 with the Unigram and TF-IDF matrix. The length feature
does not add any value to this model. The positive predicted value i.e. Precision is the
highest for the Bigram matrix having the value 1. The Recall and the F1-Score of the
Bigram model are 0.652 and 0.789 respectively which are comparatively low compared to
the Unigram and TF-IDF matrix having the values of 0.93 and 0.99 respectively.

Figure 12: Evaluation Metrics for Bernoulli Naive Bayes Model

4.11 Experiment 4: Implementation, Evaluation and Results of

Support Vector Machine Model
4.11.1 Implementation
Support Vector Machine is a supervised machine learning technique that uses an optimal
hyperplane that differentiates between 2 or more classes. (Basu et al. (2002)) SVM can
deal a large number of features with minimum error rate. The default parameters work
very well and hence explicit tuning is not required. SVM with linear kernel has shown
great results in the field of text analytics (Agarwal et al. (2016)), (Shirani-Mehr (2012)),
and hence implemented in the research as a baseline model.
Support Vector Machine classifier was implemented using Stratified 10-Fold Cross-
Validation to divide the dataset into training and test sets. The function used to im-
plement the model was the SVC() of the sklearn library. The kernel was set as linear.
The model was implemented on Unigram, Bigram, and TF-IDF with and without length
feature.
2
https://scikit-learn.org/stable/modules/classes.html

20
4.11.2 Evaluation and Results
SVM model evaluation metric is shown in the Figure 13. A comparison of all the matrices
shows that the Unigram matrix has surpassed others with an Accuracy of 0.963, Recall
of 0.945 and F1-Score of 0.962. The length feature did not contribute much to the model.
Precision for Bigram without length feature was the highest of 0.993.

Figure 13: Evaluation Metrics for Support Vector Machine Model

4.12 Experiment 5: Implementation, Evaluation and Results of

Random Forest Model
4.12.1 Implementation
Random Forest is an ensemble tree-based machine learning classifier. It is known for
generating high accuracy as mentioned in the research (Koray et al. (2019)). Random
Forest is capable of handling large datasets with high dimensionality and hence is a good
fit for the research as it handles a large number of features created by a bag of words.
Random Forest classifier was implemented using Stratified 10-Fold Cross-Validation
to divide the dataset into training and test sets. The function used to implement the
model was the RandomForestClassifier() of the sklearn library. The number of estimators
were taken as 1000. The model was implemented on Unigram, Bigram, and TF-IDF with
and without length feature.

4.12.2 Evaluation and Results

Analyzing the Figure 14, it can be observed that the length contributed to the increase in
the performance of the model for all the 3 matrices. TF-IDF with length generated the
best results in terms of Accuracy, Recall, and F1-Score. The Accuracy stood at 0.964,
Recall at 0.955 and F1-Score at 0.96. Precision for Bigram without length feature was
the highest amongst all at 0.993.

21
 Figure 14: Evaluation Metrics for Random Forest Model

4.13 Conclusion
On the basis of implementation and generated results, the research objectives till Obj-4
(section 1.3) along with the research question (section 1.2) have been accomplished. The
models developed will contribute remarkably to mobile phone users in detection of spam
SMS.

5 Discussion and Comparison of Results

5.1 Comparison of Developed Models
As can be clearly seen from the above section, the TF-IDF matrix with length produced
the best results amongst all the matrices used. Hence, the TF-IDF matrix was considered
for comparing the various models. Fig 15 and 16 illustrate the comparison of the different
models applied in this research in terms of accuracy and execution time. Evidently,
the Bernoulli Naı̈ve Bayes model generated the highest accuracy of 0.965 followed by
Random Forest and LightGBM having the accuracy of 0.964 and 0.954 respectively.
SVM and XGBoost resulted in an accuracy of 0.953 and 0.944 respectively. Random
Forest and SVM gave an accuracy that was marginally lower than Bernoulli Naı̈ve Bayes.
However, the time taken by Random Forest (34.957 seconds) and SVM (30.285 seconds)
was way higher than Bernoulli Naı̈ve Bayes which took just 0.157 seconds. The gradient
boosting models LightGBM and XGBoost took 1.708 and 7.919 seconds respectively.
The time taken by LightGBM was slightly higher than the Bernoulli Naı̈ve Bayes model
but much lower than the traditional base models like Random Forest and SVM. This
made LightGBM a good fit for the detection of spam SMS in a real-world scenario. This
attained the research objective 5 mentioned in Chapter 1(Section 1.3).

22
Figure 15: Model-wise Accuracy Figure 16: Model-wise Execution
Comparison Time Comparison

5.2 Comparison of Developed Models with Existing Models

In this section, the developed models are compared with the existing ones that are dis-
cussed in the literature review (Table 2). The researcher (Gupta et al. (2018)), achieved
an accuracy of 99.1% with the CNN model and (Kim et al. (2015)), achieved the highest
accuracy of 94.70% with Naı̈ve Bayes. As shown in the comment section of Table 3, the
reason for high and low accuracy is mentioned. The CNN model produced good res-
ults but was computationally very expensive in terms of complexity and time (Yue and
Elfayoumy (2007)).

Table 4: Comparison of Developed Models with Existing Models

References Algorithm Accuracy Comments

Gupta et al. Convolutional Neural 99.10% Researcher did not handle class
(2018) Network imbalance and hold out technique
was used to split train and test
dataset.TF-IDF matrix was used
for feature.
Kim et al. Naive Bayes 94.70% The class imbalance was not
(2015) handled. The frequency ratio
technique used for feature selec-
tion.Researcher used 10-fold Cross
Validation Technique to split test
and train dataset
Current Re- Bernoulli Naive Bayes 96.50% The class imbalance issue was
search handled, Feature selection impl-
emneted and Stratified 10-Fold
Cross Validation technique used to
split the dataset.

5.3 Critique of the Research Conducted

During the course of this research, some issues and challenges were encountered such as:
1. The SMS spam dataset used for the research was highly imbalanced. The dataset
contained 5574 records in all out of which 747 records were spam and the rest were
legitimate records. This resulted in a disparity between the spam and the ham classes

23
and the ratio was extremely uneven. This was handled by applying the down-sampling
technique to the ham class to ascertain the records were equal in ham as well as spam
class. This reduction in the record size of the ham class could have impacted the training
of the machine learning models used for this research.
2. Due to the down-sampling technique, the effective record size used for training the
machine learning models was very limited. The performance achieved using this approach
may vary when the data size grows exponentially. In the future, this can be tackled by
aggregating different datasets and training the models again for better results.
3. The research in its current form is confined to the English language only as the dataset
used contained English words. Also, the models are trained to identify proper English
language words and not the slang and short forms. The result may vary when non-English
words, slang language or short forms are used.

6 Conclusion and Future Work

Spam SMS is a grave threat and it is getting more and more serious with each day. It
can cause significant harm and the consequences can be drastic. Countering this menace
with high accuracy and low latency was the main motivation behind this research. A
sample dataset was used to find an effective solution to the above problem. In the initial
stage, Exploratory Analysis was conducted on the dataset wherein it was established
that the length feature was a contributing factor in identifying the ham and spam. This
also revealed that there was a high imbalance in the ham and spam class of the dataset.
This was taken care of by the down-sampling technique to match the ham and spam
class counts. Data pre-processing and cleaning was done to reduce the noise from the
data. Furthermore, the features were extracted using the Bag of Words and TF-IDF
models. To achieve low latency, the extracted features were selected using the Chi-Square
feature selection technique. Then the machine learning models Bernoulli Naı̈ve Bayes,
LightGBM, and XGBoost were applied along with the traditional base models SVM and
Random Forest.
The research objectives were accomplished and the spam SMS were filtered with
high accuracy within a short time. Hence the research can be termed successful. The
results section demonstrated that the suggested models like Bernoulli Naı̈ve Bayes and
LightGBM combined with TF-IDF were apt for solving the research question since they
produced an accuracy of 96.5% and 95.4% respectively. Also, the time taken by these
models was 0.157 and 1.708 seconds which was significantly better than the other tradi-
tional models.

Future Work: In an endeavor to improve on the results, the machine learning models
should be trained using datasets from different sources and also using datasets having a
large number of records. This will improve the reliability of the models. The millennial
in today’s time use slang and short forms in texting which cannot be detected by the
models at present. This can be improved upon to better classify the genuine ham messages
from the spurious ones. In-depth research can be conducted on this. Also, non-English
languages can be included for spam SMS detection in the future. Apart from Chi-Square,
other techniques like Information Gain, Gini Index, etc. can be used to evaluate the
impact on performance.

24
7 Acknowledgment
I would like to extend my heartfelt gratitude towards my mentor Dr. Catherine Mulwa.
This research would not have been possible without her guidance, help, and support. She
always went the extra mile to help me whenever I was in need. I would also like to thank
my family for bestowing their trust in me.

References
Agarwal, S., Kaur, S. and Garhwal, S. (2016). SMS spam detection for Indian messages,
Proceedings on 2015 1st International Conference on Next Generation Computing Tech-
nologies, NGCT 2015 (September): 634–638.

Aich, P., Venugopalan, M. and Gupta, D. (2019). Content based spam detection in short
text messages with emphasis on dealing with imbalanced datasets, Proceedings - 2018
4th International Conference on Computing, Communication Control and Automation,
ICCUBEA 2018 .

Akbari, F. and Sajedi, H. (2015). SMS spam detection using selected text features and
Boosting Classifiers, 2015 7th Conference on Information and Knowledge Technology,
IKT 2015 pp. 1–5.

Almeida, T. A., Almeida, J. and Yamakami, A. (2011). Spam filtering : how the dimen-
sionality reduction affects the accuracy of Naive Bayes classifiers, Journal of Internet
Services and Applications 1: 183–200.

Balli, S. and Karasoy, O. (2018). Development of content based sms classification applic-
ation by using word2vec based feature extraction, IET Software .

Basu, A., Watters, C. and Shepherd, M. (2002). Support Vector Machines for Text
Categorization, Proceedings of the 36th Hawaii International Conference on System
Sciences pp. 1–7.

Ergin, S. and Isik, S. (2014). The assessment of feature selection methods on agglutinative
language for spam email detection: A special case for Turkish, INISTA 2014 - IEEE
International Symposium on Innovations in Intelligent Systems and Applications, Pro-
ceedings pp. 122–125.

Fernandes, D., d. Costa, K. A. P., Almeida, T. A. and Papa, J. P. (2015). Sms spam
filtering through optimum-path forest-based classifiers, pp. 133–137.

Gupta, M., Bakliwal, A., Agarwal, S. and Mehndiratta, P. (2018). A Comparative Study
of Spam SMS Detection Using Machine Learning Classifiers, 2018 11th International
Conference on Contemporary Computing, IC3 2018 pp. 1–7.

Hazim, M., Anuar, N. B., Ab Razak, M. F. and Abdullah, N. A. (2018). Detecting

opinion spams through supervised boosting approach, PLoS ONE 13(6): 1–24.

Islam, M. S., Khaled, S. M., Farhan, K., Rahman, M. A. and Rahman, J. (2009). Model-
ing spammer behavior: Naı̈ve Bayes vs. artificial neural networks, 2009 International
Conference on Information and Multimedia Technology, ICIMT 2009 pp. 52–55.

25
Jindal, N. and Liu, B. (2007). Analyzing and detecting review spam, pp. 547–552.

Kim, S. E., Jo, J. T. and Choi, S. H. (2015). SMS Spam filterinig using keyword frequency
ratio, International Journal of Security and its Applications 9(1): 329–336.

Koray, O., Buber, E., Demir, O. and Diri, B. (2019). Machine learning based phishing
detection from URLs, Expert Systems With Applications 117: 345–357.
URL: https://doi.org/10.1016/j.eswa.2018.09.029

Lee, C. N., Chen, Y. R. and Tzeng, W. G. (2017). An online subject-based spam filter
using natural language features, 2017 IEEE Conference on Dependable and Secure
Computing pp. 479–484.

Li, Y., Yang, Z., Chen, X., Yuan, H. and Liu, W. (2019). A stacking model using URL and
HTML features for phishing webpage detection, Future Generation Computer Systems
94: 27–39.
URL: https://doi.org/10.1016/j.future.2018.11.004

Liew, S. W., Sani, N. F. M., Abdullah, M. T., Yaakob, R. and Sharum, M. Y. (2019). An
effective security alert mechanism for real-time phishing tweet detection on Twitter,
Computers and Security 83: 201–207.
URL: https://doi.org/10.1016/j.cose.2019.02.004

Longadge, R. and Dongre, S. (2013). Class imbalance problem in data mining review,
Int. J. Comput. Sci. Netw. 2.

Ma, J., Zhang, Y., Liu, J., Yu, K. and Wang, X. (2016). Intelligent sms spam filter-
ing using topic model, 2016 International Conference on Intelligent Networking and
Collaborative Systems (INCoS) pp. 380–383.

Minastireanu, E.-A. and Mesnita, G. (2019). Light GBM Machine Learning Algorithm to
Online Click Fraud Detection Light GBM Machine Learning Algorithm to Online Click
Fraud Detection, Journal of Information Assurance & Cybersecurity 3(April): 1–15.

Najadat, H., Abdulla, N., Abooraig, R. and Nawasrah, S. (2014). Mobile SMS Spam
Filtering based on Mixing Classifiers, International Journal of Advanced Computing
Research 1.

Pham, T.-h. (2016). Content-based Approach for Vietnamese Spam SMS Filtering, 2016
International Conference on Asian Language Processing (IALP) pp. 41–44.

Prieto, A., Prieto, B., Ortigosa, E., Ros, E., Pelayo, F., Ortega, J. and Rojas, I. (2016).
Neural networks: An overview of early research, current frameworks and new chal-
lenges, Neurocomputing 214.

Ren, Y. and Ji, D. (2017). Neural networks for deceptive opinion spam detection: An
empirical study, Information Sciences 385-386: 213–224.

Sethi, P., Bhandari, V. and Kohli, B. (2018). SMS spam detection and comparison of
various machine learning algorithms, 2017 International Conference on Computing and
Communication Technologies for Smart Nation, IC3TSN 2017 2017-October: 28–31.

26
Shafique, U. and Qaiser, H. (2014). A comparative study of data mining process mod-
els (kdd, crisp-dm and semma), International Journal of Innovation and Scientific
Research 12: 2351–8014.

Shirani-Mehr, H. (2012). SMS Spam Detection using Machine Learning Approach, tech.
rep., Stanford University pp. 1–4.

Tianqi, C. and Guestrin, C. (2016). XGBoost: A Scalable Tree Boosting System Tianqi,
Il Friuli medico 19(6).

Vani, K. and Gupta, D. (2014). Using K-means cluster based techniques in external
plagiarism detection, Proceedings of 2014 International Conference on Contemporary
Computing and Informatics, IC3I 2014 pp. 1268–1273.

Xu, S. (2018). Bayesian naı̈ve bayes classifiers to text classification, Journal of Informa-
tion Science 44(1): 48–59.
URL: https://doi.org/10.1177/0165551516677946

Yadav, K., Saha, S. K., Kumaraguru, P. and Kumra, R. (2012). Take control of your
SMSes: Designing an usable spam SMS filtering system, Proceedings - 2012 IEEE 13th
International Conference on Mobile Data Management, MDM 2012 pp. 352–355.

Yu, B. and ben Xu, Z. (2008). A comparative study for content-based dynamic
spam classification using four machine learning algorithms, Knowledge-Based Systems
21(4): 355–362.

Yuan, H., Chen, X., Li, Y., Yang, Z. and Gv Liu, W. (2018). Detecting Phishing Web-
sites and Targets Based on URLs and Webpage Links, Proceedings - International
Conference on Pattern Recognition 2018-August: 3669–3674.

Yue, Y. and Elfayoumy, S. (2007). Anti-spam filtering using neural networks and Baysian
classifiers, Proceedings of the 2007 IEEE International Symposium on Computational
Intelligence in Robotics and Automation, CIRA 2007 pp. 272–278.

27