EMAIL SPAM DETECTION USING HYBRID ALGORITHM

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 10 Issue: 09 | Sep 2023 www.irjet.net p-ISSN: 2395-0072
© 2023, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 628
EMAIL SPAM DETECTION USING HYBRID ALGORITHM
RAGAVI R 1, JANANI G2,
1Student, Dept. of Information Technology, Bannari Amman Institute of Technology, Tamil Nadu, India
2 Student, Dept. of Information Technology, Bannari Amman Institute of Technology, Tamil Nadu, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Email spam is still a continuous and changing
issue that threatens user experience and calls into question
the reliability of digital communication. By automating the
detection of spam emails, machine learning techniques have
become a potent weapon in the fight against this problem.
In this study, a hybrid algorithm that combines both Logistic
Regression and Neural Networks for email spam detection is
introduced. The simplicity of Logistic Regression and
pattern recognition capabilities of neural networks have
been combined. Also, to prepare the data for the model,
feature extraction and data transformation is involved. At
first, Logistic Regression identifies the spam indicators in
the mail which is followed by deeper analysis by Neural
Networks to enhance classification. The hybrid algorithm
overtakes individual models, according to experimental
findings, and achieves a high level of accuracy in
differentiating spam from authentic emails leading to have
a more secure digital communication environment.
Key Words: machine learning, Neural Networks, Secure
Digital Communication, Logistic Regression.
1. INTRODUCTION
In the present technological world, the widespread use of
email communication has completely changed how we
interact and share our information both official and
personal. The actual problem here is, as the usage of email
grows immensely, Spam, phishing and all other malicious
activities also grow along with it. The immense increase in
the number of unwanted emails called spam and the
person sending them are spammers. Online spam has
grown to be a serious issue. It blocks the user from making
good usage of storage and time. Also, the massive amount
of spam mails flourishing through the network have
devastative effects on the memory space of mail servers,
CPU power and communication bandwidth. The level of
danger of spam emails is on the increase in daily basis and
is responsible for over 77% of global email traffic. These
spam emails have resulted in financial loss to many users
who were the victim of fraudulent practices of spammers
who send emails pretending to be from reputed
companies with the aim to gain sensitive information like
passwords, credit card numbers etc., These has resulted in
the need for development of the model “spam detection
and filtering”. This proposed hybrid machine learning
algorithms provide a proactive and reliable way to
successfully combat these risks.
Email spam is still a problem, and fraudsters' strategies
are getting more and more complex. Regular spam filters
frequently find it difficult to keep up with these developing
dangers. In order to achieve greater accuracy in
identifying and blocking harmful emails, our solution
acknowledges the need for a more flexible and robust
approach, integrating the strengths of several machine
learning algorithms. The leading email providers such as
Gmail and Yahoo have employed different machine
learning models such as Naïve Bayes, Neural networks,
Support vector machine and K Nearest neighbor. We are
thrilled to present our cutting-edge method for email
spam identification, utilizing the strength of hybrid
machine learning algorithms, to battle this ever-evolving
threat. In this project, the model has been made with the
use of hybrid algorithms, which combine the strengths of
different machine learning models like neural networks,
decision trees, and support vector machines. These
algorithms can be combined to more effectively capture
the many patterns and attributes found in email content,
headers, sender behavior, and attachment properties.
2. LITERATURE SURVEY
Kriti Agarwal, Tarun Kumar [2020] makes an effort to
apply machine learning techniques to recognize a pattern
of recurring keywords that are categorized as spam. The
system also suggests categorizing emails based on
additional variables found in their structure, such as
Cc/Bcc, domain, and header. When a parameter is applied
to the machine learning algorithm, it is regarded as a
feature. The pre- trained machine learning model will have
a feedback mechanism to discriminate between a correct
output and an ambiguous output. This approach offers an
alternate architecture for the implementation of a spam
filter.
Aakash Atul Alurkar [2019] examined the usage of string
matching algorithms for identifying spam emails. The
performance of six well-known string matching
algorithms, including the Longest Common Subsequence
(LCS), Levenshtein Distance (LD), Jaro, Jaro-Winkler, Bi-
gram, and TFIDF, is specifically examined and compared in
this paper using the Enron corpus and the CSDMC2010
spam dataset as two different datasets. They found that in
both datasets, the Bi-gram algorithm performs the best at
detecting spam.

T.Verma [2018] proposed a technique for filtering emails
using the SVM algorithm and feature extraction. This
method includes a number of processes, including Email
Collection, where data is taken from the dataset. It is then
routed via preprocessing, wherein extraneous content is
eliminated and only desired content is sent on to the next
step. Then comes feature extraction followed by SVM
model training. The dataset from the Apache Public
Corpus was used by the author. Special symbols, HTML
tags, URLs, and extraneous alphabets were deleted from
the suggested solution. All of the dictionary words were
mapped by the author using the vocabulary file. A 98%
accuracy was achieved using the SVM method on a pre-
processed dataset.
Abdulhamid Muhammad Shafi [2018] conducted
performance analyses for various machine learning
classification methods, including Random Forest, Lazy
Bayesian Rule, and Radial Basic Function (RBF) Network.
Bayesian Logistic Regression, J48, and Tree. There was a
based on a comparison of all of these provided algorithms
the F-measure, precision, recall, root mean squared error,
and Accuracy. They made use of the UCI Machine dataset.
Learning Archive. For determining the recall and precision
They used the F-measure approach to determine value.
Rotation forest algorithm was used to acquire the
maximum measure, and Naive Bayes algorithm has the
lowest. The best result was reached for the Rotation
Forest Algorithm with 87.9 using the Kappa Statistics for
the statistical results. Rotation Forest technique produced
the highest accuracy (94%), while REP Tree algorithm
produced the lowest accuracy (89%). Other algorithms,
such Naive Bayes and J48, provided accuracy of 88.5% and
92.3%, respectively.
3. OBJECTIVE AND METHODOLOGY
This project mainly focuses on developing an efficient
email spam classifier which can identify and filter emails
more accurately than ever. It prevents spam messages
from getting loaded into the user' inbox, thereby
improving user experience. It is known that nowadays
emails are just flourishing in the inbox which consists of
both spam and ham. These ham emails are very dangerous
because of the fraudulent practices of spammers who sent
these kinds of illegitimate emails. Because of this, users
have even suffered from financial loss. This is considered
to be one of the reasons behind our objectives. Along with
this, spam emails also occupy most of the storage space. It
also wastes the time because without spam detection and
filtering, the user has to do it on their own.
Compared to the existing algorithms, this model is
designed to be more scalable and can handle large
volumes of email data. The hybrid approach which
combines both neural networks and logistic regression
can add the strengths of both the techniques. While
logistic regression offers interpretability and can act as a
useful final classifier, neural networks are powerful for
feature learning and catching complicated patterns. This
model supports false positives reduction.
Even these days, many spam emails are incorrectly
classified as ham and ham emails are incorrectly classified
as spam which creates a lot of confusion for the user. This
issue is known as false positives. This model overcomes
the above issue as the algorithm used here is neural
networks. The model can generalize good to new spam
patterns and those unseen ones without getting the
training data overfitted. It increases the overall user
experience by reducing its exposure to spam emails in the
inbox.
4. CONTRIBUTION
The contribution for the successful completion of the
model has been broken into two parts where my part is
completely focused into training and testing stages. In the
context of machine learning, these two stages are
considered very crucial for making sure of the model’s
effectiveness and accuracy in serving its calculated
functions. In order to begin with I have started to
thoroughly research all of the different algorithms that
were relevant to this project. Because, only with proper
research, it is possible to choose the most suitable
algorithm.
After understanding all the algorithms, each was put into
implementation for getting the accuracy rate for further
proceedings in the project which was considered to be a
baseline for our endeavour. At first the model was
individually trained with each algorithm. After getting
results, it soon became clear that using just one algorithm
wouldn't be enough to produce the necessary results. As a
result, we made the decision to research the idea of hybrid
algorithms. In an effort to achieve greater accuracy, we
tried mixing various algorithms.
During the individual evaluation of each algorithm, it is
known that logistic regression and k nearest neighbour
algorithm has shown greater accuracy as compared with
the others. But the thing is, all just work upon the
frequency of occurrence of each word. We were aware of
the limits of these conventional algorithms as our research
developed. They had some effectiveness, but they were
unable to understand the larger context of text messages.
Instead of using these algorithms, it is good to switch to
the concept of neural networks where the model will be
reading the whole text message, understands the context
and semantics and then proceed with further jobs. This
capacity was expected to considerably improve the
precision and potency of our model. From all the above
findings, it was decided that it is better to use the

combination of both neural networks and logistic
regression.
5. FLOW DIAGRAM
EXPLANATION OF FLOW DIAGRAM:
In this step, the email data will be prepared so that the
machine learning models may be trained and tested.
• Lowercasing: we have converted the uppercase letter to
lowercase letters to make sure that the model should not
treat them different as both are same.
• Removal of special characters and punctuations: As
punctions and special characters like ‘$’,’&’ etc., don’t carry
any necessary information, we have removed them in
order to make the dataset compact.
• Handling newlines and blank spaces(whitespaces): In
addition to the above process, we have also removed extra
newlines to support text formatting.
• Tokenization: we had split the text document into a list
of words that is known to be tokens. Each token in the text
serves as a unique unit of meaning. We consider this step
as very important because it enables you to treat each
word as a separate object, enabling fine-grained analysis
and processing of the text. The machine learning model
uses each token as a feature in order to make predictions.
• Removal of Stop Words: In this step, we have removed
the words that occur very frequently in text and also
doesn’t carry any necessary information in it. Performing
this step help in dimensionality reduction which improves
the model performance further. In English, the words that
frequently occur are ‘the’, ’and’, ’or’, ’in’, ’is’ etc.,
•Stemming and Lemmatization: Lemmatization and
stemming are methods for breaking down words to their
root or basic forms. The main objective is to combine
words with comparable meanings, which can aid in
feature engineering and feature reduction. When you wish
to minimize dimensionality but don't necessarily need
proper words, stemming is a good approach. But in the
case of lemmatization, it is a more complex method that
reduces words to their dictionary form. The outcome is
always a meaningful word. When you wish to keep the
words' grammatical correctness and interpretability,
lemmatization is preferred.
• Feature Extraction: we have created numerical features
by converting the preprocessed text data. These numerical
features are considered to be the inputs for our machine
learning model. This technique includes term TF- IDF
which we have implemented. It's a numerical
statistic that is frequently used in information
retrieval and natural language processing to assess
the significance of a word inside a document. For
text-based machine learning applications which
includes text categorization and information retrieval, TF-
IDF is especially helpful. The frequency of a particular
word in a document is quantified by the TF.
• Data Preparation for Neural Networks: As we used
neural networks, we made all the email texts to be in same
length by truncating larger texts or padding shorter ones
with zeros. This is known as padding sequences. Apart
from this, we replaced the spam and legitimate labels into
numerical values that is, replacing ‘0’ for legitimate or ham
labels and ‘1’ for spam labels. This process is known as
encoding labels.
• Data Preparation for Logistic Regression: As we have
used logistic regression for training our model, we used
the preprocessed features like the tokens, TF-IDF vectors
from which it classifies accordingly. When a new mail
arrives after all the preprocessing features have been
generated, it calculates the frequency of the words and
then decides whether these words occurred mostly in
spam or ham emails from which it has been trained. And
then, it classifies as spam or ham.
• Data Splitting: We have splitted the preprocessed dataset
into training and testing data in a ratio of 80:20. The
machine learning model is trained using the data from the
subset ‘training set’. This part of the data is used by the
model to learn patterns and relationships. The

performance of the final model is assessed using the
testing set. It is a chunk of the data that has been withheld
and has not been viewed by the model during training or
validation. Apart from training and testing set, we have
validation set. This set helps us in selecting the good
model among others. It evaluates the performance of the
model at the time of training. Also, it prevents overfitting.
It is good to randomize the rank of the samples to make
sure that the sets reflect a vast number of examples. Also,
in order to maintain the class distribution in each split, we
have utilized stratified sampling. This made sure that
each set accurately represents the proportions of the
entire class.
6.RESULT AND DISCUSSION
A confusion matrix is a summary of the results of a
categorization challenge. The number of right and bad
predictions are totalled with count values and dispersed
by each class. The confusion matrix's key is this. The
confusion matrix shows how your classification model
becomes perplexed while making predictions. It gives us
insight into the types of errors that are being made, as well
as into the mistakes that specific people have made when
using a classifier. Only after the true values of the test data
are known can it be determined.
Here,
Set 1: Positive
Set 2: Negative
● TP: Examination will be positive, also it’s predicted
value will be positive.
● FN: Examination will be positive, but it’s predicted value
will be negative.
● TN: Examination will be negative, but it’s predicted
value will be negative.
● FP: Examination is negative, but it’s predicted value will
be positive
6.1. NAIVE BAYES
Naive Bayes is a popular machine learning method for
email spam identification. Since Naive Bayes is a simple
understandable algorithm, it has a short implementation
time. When you just have a tiny amount of labelled data,
Naive Bayes is advantageous since it performs well even
with small training datasets. Using Bayes' theorem, we
have calculated the posterior probabilities for both the
spam and non-spam classes in order to categorize a new
email as spam or not. In conclusion, Naive Bayes also have
drawbacks, its naive independence assumption and
constrained expressiveness. It can be a good option for
simple spam detection jobs, but other algorithms, such as
logistic regression, neural networks may produce much
more good results.
6.2. LOGISTIC REGRESSION
Logistic Regression is another algorithm that can be used
for email spam detection in machine learning. Using the
logistic (sigmoid) function, logistic regression models the
probability that an email belongs to the spam class. In
order to maximize the likelihood of the observed data, the
logistic regression algorithm learns the coefficients for
each feature during training. By penalizing large
coefficients, regularization approaches like L1 or L2
regularization can be used to prevent overfitting. The
probabilities generated by logistic regression fall between
[0, 1]. To choose the class label, we selected a threshold
almost like 0.5. The precision and recall trade-off can be
adjusted by changing the threshold. Results from logistic
regression are simple to interpret. It is a reliable option
when working with noisy data or when you have a limited
amount of labelled data because it has a decreased risk of
overfitting compared to more complex algorithms. It also
works on the frequency of the words present.

6.3. HYBRID ALGORITHM
Combining Logistic Regression and Neural Networks in
email spam detection can offer several advantages, often
referred to as a hybrid approach. Performance can be
enhanced by combining various techniques, such as neural
networks and logistic regression. This hybrid technique
efficiently captures both simple and complicated patterns
by combining the advantages of both models. We fed the
neural network's newly learned features into a logistic
regression model after it has been trained. A set of input
features are used in a simple linear model called logistic
regression to provide an output that may be understood as
the likelihood that an email is spam. For binary
classification problems like email spam detection, it is a
widely used model. you use the neural network's output as
a set of logistic regression's input features. Logistic
regression uses the features that the neural network has
extracted as its input features. The following is the
accuracy we obtained from our hybrid model. Hybrid
algorithms play a prominent role in improving the search
capability of algorithms. Hybrid algorithms exploit the
good properties of different methods by applying them to
problems they can efficiently solve. The fusion of more
than one algorithm makes the model to effectively classify
the email thus making it the best one.
7. CONCLUSION
In conclusion, the hybrid strategy that combines logistic
regression and neural networks for email spam detection
offers a complete and efficient solution. The ability of
neural networks to recognize complex patterns in email
content and the clarity and simplicity of logistic regression
are two advantages of this combination. The resulting
hybrid model is perfect for real-time email identification
because it not only boasts improved accuracy but also
feature extraction abilities and scalability. Its robustness is
highlighted by its capacity to adjust to various spam forms
and shifting strategies. However, to ensure user privacy
and uphold legal requirements, the successful
implementation of such a system calls for rigorous model
tuning, continual monitoring, and ethical considerations.
By achieving these goals, the hybrid model guarantees
customers a dependable and seamless email experience,
effectively eliminating spam while upholding openness
and credibility throughout.
8. SUGGESTIONS FOR FUTURE WORK
To further improve the efficiency of email spam detection
system, there are some suggestions which provide
directions for achieving the above. Using advanced deep
learning architectures such as BERT model for email spam
detection can be said as a good approach. It is a
transformer-based model. Other than its much higher
computational resources, it can be considered as a very
good approach.
Other than this, we can also include integration of user
feedback. It integrates feedback on the emails that are mis
classified in order to improve the email spam detection
and filtering system performance.
REFERENCES
[1] T. A. Almeida and A. Yamakamil, “Facing the spammers:
A very effective approach to avoid junk Emails,” Expert
Systems with Applications, vol. 39, Issue 7, 1 June 2012,
Pages 6557–6561.
[2] T. Fawcett. “An introduction to ROC analysis,” Pattern
Recognition Letters, vol. 27, issue 8, pp. 861-874, June
2006.
[3] T. S. Guzella and W. M. Caminhas, “A review of machine
learning approaches to spam filtering,” Expert Systems
with Applications, vol. 36, pp. 10206–10222, 2009.
[4] T. Guzella and W. Caminhas, “A review of machine
learning approaches to spam filtering,” Expert Syst. Appl.,
vol. 36, no. 7, pp. 10206–10222, 2009.
[5] Q. Le and T. Mikolov, “Distributed representation of
sentences and documents,” presented at the 31 the

International Conference on Machine Learning, Beijing,
China, 2014.
[6] M. Woitaszek, M. Shaaban, and Czernikowski.
“Identifying junk electronic mail in Microsoft outlook with
a support vector machine,” in Proc. Symposium on
Applications and the Internet, 2003, pp. 66–169.
[7] Q. Le and T. Mikolov, “Distributed representations of
sentences and documents,” arXiv:1405.4053.
[8] J. Hovold, “Naive Bayes spam filtering using word-
position-based attributes,” presented at 2nd Conference
on Email and Anti-Spam, Stanford, CA, 2005.
[9] I. Kanaris, K. Houvardas, and E. I. Stamatatos, “Words
vs. Character n-grams for anti-spam filtering,”
International Journal on Artificial Intelligence Tools, vol.
16, no. 6, pp. 1047–1067, 2007.
[10] G. Salton, Introduction to Modern Information
Retrieval, A. G. K. Janecek, W. N. Gansterer, M. A. Demel,
and G. F. Ecker, “On the relationship between feature
selection and classification accuracy,” in Proc. the
Workshop on New Challenges for Feature Selection in
Data Mining and Knowledge Discovery, 2008, p.
16.Auckland, McGraw-Hill, 1983.
[11] J. S. Lee and D.-W. Kim, “Fast multi-label feature
selection based on information-theoretic feature ranking,”
Pattern Recognition, vol. 48, issue 9, pp. 2761-2771, 2015.
[12] B. Trstenjak, S. Mikac, and D. Donko, “KNN with TF-
IDF based Framework for Text Categorization,” Procedia
Engineering, vol. 69, pp. 1356-1364, 2014.
[13] M. Sahami, S. Dumaisis, D. Heckerman, and E. Horvitz,
“A Bayesian approach to filtering junk e-mail,” AAAI
Technical Report WS-98-05, 1998.
[14] G. E. Hinton, “Learning distributed representations of
concepts,” inProc. the Eighth Annual Conference of the
Cognitive Science Society,”
[15] D. Talbot, “Where SPAM is born,” MIT Technol. Rev.,
vol. 111, no. 3, p. 28, 2008.
[16] C. Li, L. Ji, and J. Yan. “Acronym disambiguation using
word embedding,” in Proc. the Twenty-Ninth AAAI
Conference on Artificial Intelligence, 2015.
[17] S. Bengio and G. Heigold, “Word embeddings for
speech recognition,” in Proc. 15th Annual Conference of
the International Speech Communication Association,
Singapore, September 14-18, 2014.
[18] A. Barushka et al., “Spam filtering using regularized
neural networks with rectified linear uni.,” presented at
XVth International Conference of the Italian Association
for Artificial Intelligence, Genova, Italy, 2016.
[19] Yuchun Tang, Sven Krasser, Yuanchen He, Weilai
Yang, Dmitri Alperovitch ”Support Vector Machines and
Random Forests Modeling for Spam Senders Behavior
Analysis” IEEE GLOBECOM, 2008
[20] Steve Webb, James Caverlee, Calton Pu, Introducing
the Webb Spam Corpus: using Email spam to identify web
spam automatically, CEAS, 2016.

EMAIL SPAM DETECTION USING HYBRID ALGORITHM

More Related Content

Similar to EMAIL SPAM DETECTION USING HYBRID ALGORITHM

More from IRJET Journal

Recently uploaded

EMAIL SPAM DETECTION USING HYBRID ALGORITHM