Arabic sentiment analysis of Monkeypox using deep neural network and optimized hyperparameters of machine learning algorithms

Abstract

Sentiment analysis, a branch of natural language processing (NLP), has gained significant attention for its applications in various domains. This study focuses on utilizing machine learning and deep learning algorithms for sentiment analysis in the context of analyzing Monkeypox using Arabic sentiment text. The objective is to develop an accurate and efficient model capable of classifying Arabic text into sentiment categories, facilitating the understanding of public perceptions toward Monkeypox. The study begins by collecting a diverse dataset of Arabic text containing sentiments related to Monkeypox. Machine learning algorithms, such as Support Vector Machines, Naive Bayes, and Random Forest, along with deep learning (DNN) techniques, including Recurrent Neural Networks and Transformer models, are employed for sentiment classification. Hyperparameter optimization techniques were implemented to fine-tune the models for optimal performance. The impact of various hyperparameters on the model is assessed to select the best configuration. Experimental results demonstrate the effectiveness of the proposed sentiment analysis models in accurately classifying Arabic sentiment text related to Monkeypox. The DNN models based on Leaky ReLU showcased the significance of leveraging complex representations for NLP tasks with 92%. Hyperparameter optimization aids in selecting suitable configurations, improving model accuracy, and reducing overfitting. The findings from this study contribute to advancing sentiment analysis techniques in Arabic text and provide valuable insights into public sentiments toward Monkeypox. The developed models can be utilized in public health monitoring, crisis management, and policymaking, offering valuable insights into the sentiment landscape surrounding the disease.

Full text

Vol.:(0123456789)
Social Network Analysis and Mining (2024) 14:30
https://doi.org/10.1007/s13278-023-01188-4
ORIGINAL ARTICLE
Arabic sentiment analysis ofMonkeypox using deep neural network
andoptimized hyperparameters ofmachine learning algorithms
HasanGharaibeh1· RabiaEmhamedAlMamlook2,3· GhassanSamara4· AhmadNasayreh1· SajaSmadi1·
KhalidM.O.Nahar1· MohammadAljaidi4· EssamAl‑Daoud4· MohammadGharaibeh5· LaithAbualigah6,7,8,9,10,11,12
Received: 18 October 2023 / Accepted: 18 December 2023
© The Author(s), under exclusive licence to Springer-Verlag GmbH Austria, part of Springer Nature 2024
Abstract
Sentiment analysis, a branch of natural language processing (NLP), has gained significant attention for its applications in
various domains. This study focuses on utilizing machine learning and deep learning algorithms for sentiment analysis in
the context of analyzing Monkeypox using Arabic sentiment text. The objective is to develop an accurate and efficient model
capable of classifying Arabic text into sentiment categories, facilitating the understanding of public perceptions toward
Monkeypox. The study begins by collecting a diverse dataset of Arabic text containing sentiments related to Monkeypox.
Machine learning algorithms, such as Support Vector Machines, Naive Bayes, and Random Forest, along with deep learning
(DNN) techniques, including Recurrent Neural Networks and Transformer models, are employed for sentiment classification.
Hyperparameter optimization techniques were implemented to fine-tune the models for optimal performance. The impact
of various hyperparameters on the model is assessed to select the best configuration. Experimental results demonstrate the
effectiveness of the proposed sentiment analysis models in accurately classifying Arabic sentiment text related to Monkeypox.
The DNN models based on Leaky ReLU showcased the significance of leveraging complex representations for NLP tasks
with 92%. Hyperparameter optimization aids in selecting suitable configurations, improving model accuracy, and reducing
overfitting. The findings from this study contribute to advancing sentiment analysis techniques in Arabic text and provide
valuable insights into public sentiments toward Monkeypox. The developed models can be utilized in public health monitor-
ing, crisis management, and policymaking, offering valuable insights into the sentiment landscape surrounding the disease.
Keywords Arabic sentiment analysis· Machine learning· Optimization· Hyperparameter tuning· DNN· NLB
1 Introduction
Sentiment analysis is considered one of the important topics
in recent years, as it analyzes people’s opinions about a spe-
cific problem or topic, whether the opinion is negative, posi-
tive, or neutral. Information is gathered from social media
platforms such as Facebook, Twitter, and YouTube (Cambria
2022). Natural language processing plays a significant role
in this field, whether it is removing stop words, performing
stemming, and much more. Sentiment analysis targets vari-
ous fields, including politics, economics, social issues, and
health. This approach allows companies to assess customer
satisfaction with their products or services. This complex-
ity of Arabic includes multiple dialects and morphological
aspects (El-Beltagy and Ali 2013).
One area where sentiment analysis can have a significant
impact is in health care and disease monitoring. Accurate
and timely identification of sentiments related to specific
diseases can help public health authorities and medical pro-
fessionals understand public perception, identify potential
outbreaks, and respond effectively. In this context, Monk-
eypox, a viral disease that affects humans and animals, pre-
sents a unique challenge. Monkeypox outbreaks can lead to
panic, misinformation, and fear among the public. Monitor-
ing sentiment related to Monkeypox can provide early warn-
ing signs, identify areas of concern, and guide public health
interventions. In recent times, deep learning has been used
in sentiment analysis for Arabic language databases. The
number of studies that have utilized deep learning for Arabic
language research is very limited. The database was col-
lected from Twitter for this study, given the increased inter-
action on social media platforms. These platforms serve as
a space for expressing opinions and exchanging information
Extended author information available on the last page of the article

Social Network Analysis and Mining (2024) 14:30 30 Page 2 of 18
on various aspects of life. According to a study conducted
by Arab social media outlets, the number of Arabic users
on Twitter reached 11 million, with the number of tweets
exceeding 849 million tweets (Salem 2017).
This paper investigates the problem of Monkeypox by
proposing a new sentiment analysis approach using machine
learning and deep learning to identify the features that play
a significant role in capturing sentiment from tweets. The
tweets written in Arabic were classified into three catego-
ries: positive, neutral, and negative. The following models
were used to conduct this study: Support Vector Machine
(SVM), AdaBoost, XGBoost, and LGBM (Light Gradient
Boosting Machine). By creating a robust sentiment analysis
methodology specifically designed for Arabic sentiment text
related to Monkeypox, we can gain a valuable understand-
ing of public sentiment and swiftly address concerns while
countering the spread of false information. The results of
this study have the power to significantly contribute to public
health initiatives by equipping decision-makers with crucial
tools for effective disease management and control.
The paper is organized as follows: Sect.2 contains the
related previous works on sentiment analysis in the Ara-
bic language. Section3 includes a detailed explanation of
the database, along with its preprocessing. It also describes
the models used in this study. Section4 presents the results
obtained after applying the proposed approach to the
database.
2 Related work
Researchers have conducted extensive studies on sentiment
analysis and disease monitoring, covering various aspects
and employing different techniques. A study by Atoum and
Nouman (2019) proposed a model for categorizing tweets
in the Jordanian dialect into negative, neutral, and posi-
tive classes. They utilized Naive Bayes and Support Vector
Machine (SVM) algorithms, with SVM achieving an accu-
racy of 82.1%. The study collected 1000 tweets from Twitter
using a dedicated reading application.
Tan etal. (2023) proposed a novel model for sentiment
analysis. They applied emotional variance analysis on stu-
dent journals garnered from an experiential course, and
they found that EVA is helpful for profiling variations in
sentiment polarity, with results showing an accuracy of
88.7% using a Multi-Layer Perceptron (MLP) machine
learning model. Jain etal. (2023) presented two methods
to visualize the results in order to interpret the system’s
decisions. The study used sentiment analysis based on
Natural Language Processing. Valence Aware Diction-
ary for Sentiment Reasoning (“VADER”) and Locally
Interpretable Model-Agnostic Explanations (“LIME”) have
been used for visual justification of the result, increasing the
understanding.
Rukhsar etal. (2023) performed experiments on a dataset
of 90,000 tweets relevant to the COVID-19 pandemic using
both deep learning and machine learning methods in order
to understand the psychological impact of the pandemic on
people. The deep learning model Long Short-Term Memory
(LSTM) and the Support Vector Machine (SVM) classifier
both achieved 90% accuracy. Huang etal. (2023) detected
the techniques used for sentiment analysis (SA) in current
e-commerce platforms and the future directions for e-com-
merce. The paper has chosen 54 experimental papers for
review, 26 of which have employed machine learning tech-
niques. In contrast, 24 employed SA through deep learning
techniques, and four employed both machine learning and
deep learning techniques.
Rodríguez-Ibánez et al. (2023) conducted a
comprehensive review of the multifaceted reality of
sentiment analysis in social networks about any topic on
this platform. This paper reviews the domains where
these techniques have been applied, including academic
perspective, causal relationships, temporal dynamics, and
applications in industry. Heikal etal. (2018) focused on
deep learning approaches for sentiment analysis of Arabic
tweets. They employed the Long Short-Term Memory
(LSTM) and Convolutional Neural Network (CNN) models
to predict sentiment. Their study utilized the Arabic
sentiment tweets dataset (ASTD), which contained 10,000
tweets categorized into negative, neutral, positive, and
objective sentiments. The models achieved an F1-score
of 64.46%. Sayed etal. (2020) conducted a classification
task on a database of hotel reviews from the Booking.com
website. The database consisted of 6318 reviews written in
various forms of the Arabic language. The study utilized
nine models, including Naive Bayes, K-Nearest Neighbor
(KNN), Multi-layer Perceptron (MLP), Random Forest
(RF), Support Vector Machine (SVM), Decision Tree
(DT), Gradient Boosting (GB), Ridge Classifier (RC), and
Logistic Regression (LR). The Ridge Classifier attained the
highest accuracy of 95.21%. In their study, Alayba etal.
(2018) introduced a sentiment analysis approach for Arabic
text using Convolutional Neural Network (CNN) and Long
Short-Term Memory (LSTM) models. Multiple databases
were utilized, including datasets related to Arabic health-
care services, Twitter, and Arabic sentiment tweets. The
study achieved impressive accuracies of 94.24%, 95.68%,
and 92% for each respective database. Oussous etal. (2020)
proposed an approach employing Convolutional Neural
Network (CNN), Long Short-Term Memory (LSTM),
Naive Bayes, and Support Vector Machine (SVM) models
for sentiment analysis. Various preprocessing techniques,
such as normalization, stop-word removal, stemming,

Social Network Analysis and Mining (2024) 14:30 Page 3 of 18 30
and tokenization, were applied. The Moroccan Sentiment
Analysis Corpus (MSAC), comprising 2000 reviews, was
used for evaluation, resulting in an accuracy of 96%. Hadwan
etal. (2022) presented an approach to determine user
opinions on different applications in the Kingdom of Saudi
Arabia. The researchers collected a database of 8000 reviews
from social media platforms and Google Play, which was
subsequently reduced to 7759 reviews after preprocessing.
Multiple classifiers, including Naive Bayes, Decision Tree,
Support Vector Machine (SVM), and K-Nearest Neighbor
(KNN), were employed. The approach achieved an accuracy
of 78.46%.
Baker etal. (2020) conducted a study on the sentiment
classification of Arabic tweets related to influenza. Multi-
ple classifiers, such as Decision Tree (DT), Support Vec-
tor Machine (SVM), Naive Bayes, and K-Nearest Neighbor
(KNN), were utilized. The dataset consisted of 6300 tweets
collected from Twitter. The Naive Bayes classifier achieved
the highest accuracy of 89.06%. Gamal etal. (2019) focused
on analyzing emotions in various Arabic dialects. They
gathered a large collection of 151,000 tweets and classified
them into positive and negative sentiments. Support Vec-
tor Machine (SVM), Naive Bayes, Ridge Regression, and
AdaBoost models were applied, with the Ridge Regression
model achieving an impressive accuracy of 99.9%. In their
study, Mohammed and Kora (2019) proposed an advanced
approach utilizing deep learning techniques to analyze emo-
tions across multiple topics. They employed a dataset com-
prising 40,000 tweets collected from Twitter. Three classi-
fiers, namely, Convolutional Neural Network (CNN), Long
Short-Term Memory (LSTM), and RCNN, were utilized,
along with the utilization of word embeddings. The LSTM
model demonstrated the highest accuracy of 81.31%. Nota-
bly, the study observed an 8.3% increase in accuracy when
increasing the dataset size. Aloqaily etal. (2020) focused
on sentiment analysis in the context of the Syrian crisis
and civil wars using the Arabic tweets dataset. This data-
set consisted of 2000 tweets, and various machine learn-
ing algorithms, including Simple Logistic, Logistic Model
Trees (LMT), Support Vector Machine (SVM), K-Nearest
Neighbors (KNN), and Vote, were employed. The proposed
approach achieved commendable results, including an accu-
racy of 85.55%, an AUC of 86%, and an F1-score of 92%.
Hnaif etal. (2021) proposed an approach for sentiment
analysis, incorporating normalization, stemming, and stop-
word removal techniques. They utilized classifiers such as
Naive Bayes, Support Vector Machine (SVM), K-Nearest
Neighbor (KNN), Random Forest, and Decision Tree. The
database consisted of 6138 tweets and posts collected from
Facebook and Twitter at both local and international levels.
The Support Vector Machine achieved the highest F1-score,
reaching 83%.
In (Al-Tamimi etal. 2017), supervised models were
employed to classify comments on YouTube. The research-
ers collected a database from YouTube encompassing 8053
comments from 23 Arab countries. They utilized models
including K-Nearest Neighbor (KNN), Bernoulli Naive
Bayes, and Support Vector Machine with RBF kernel,
achieving an accuracy of 88.8%. Abu-Farha and Magdy
(2020) focused on the detection of sarcastic tweets using the
ArSarcasm database, which contained 10,547 tweets with
16% of them being sarcastic. They employed a model called
BiLSTM, which achieved an accuracy of 46%. Alwakid etal.
(2017) analyzed Arabic sentiments related to unemployment
in the Kingdom of Saudi Arabia using a dataset of 4000
tweets. They utilized Support Vector Machine (SVM) and
Naive Bayes classifiers, achieving an accuracy of 73%. In
the study by Alayba etal. (2017), tweets regarding Arab
opinions on health-care services were collected from Twit-
ter, resulting in a filtered database of 2026 tweets. Various
machine learning models, including Logistic Regression,
Support Vector Machine (SVM), Naive Bayes, Stochas-
tic Gradient Descent, and Convolutional Neural Network
(CNN), were employed. The study achieved an accuracy of
90.14%. Table1 shows the previous studies closely to pro-
posed work.
The previous studies have made valuable contributions
to the field of sentiment analysis in different contexts. How-
ever, they also come with certain gaps and limitations that
need to be addressed. One common limitation observed
in some studies is the scarcity of data, which can impact
the accuracy of the sentiment analysis approach. This
emphasizes the need for larger and more diverse datasets
to improve the robustness of the models. Furthermore, a
few studies encountered challenges with the suitability of
the applied methods for the specific type of data they were
analyzing. It is crucial to select appropriate techniques that
align with the characteristics and nuances of the given data-
set to ensure accurate sentiment analysis results. While the
mentioned studies have explored sentiment analysis in vari-
ous domains, such as health care and social media, there is
a gap in the literature when it comes to studying sentiment
analysis specifically for Monkeypox using Arabic sentiment
text. Due to the lack of research interested in the Arabic
language, in addition to the fact that Monkeypox is one of
the diseases that have gained wide popularity in terms of
its cause and source, different opinions have emerged about
this disease, some of which were negative and some were
positive. This highlights an opportunity for future research to
fill this gap and develop a comprehensive sentiment analysis
approach tailored to analyze Arabic sentiment text related
to Monkeypox.
This study will improve disease monitoring. By develop-
ing a robust sentiment analysis approach for Monkeypox
using Arabic sentiment text, this enables early detection

Social Network Analysis and Mining (2024) 14:30 30 Page 4 of 18
of potential outbreaks, identification of areas of concern,
and proactive intervention strategies. Therefore, this study
will be insights into public sentiment. This study will offer
valuable insights into public sentiment toward Monkeypox
through sentiment analysis of Arabic text. Understanding
public opinions, emotions, and attitudes toward the disease
is crucial for effective public health campaigns and com-
munication strategies. The insights gained from this work
can guide targeted interventions, address concerns, and dis-
pel misinformation to foster trust and cooperation within
affected communities. Finally, this study will be as decision-
making tool for disease management. The developed senti-
ment analysis approach will provide decision-making tools
for public health professionals in the field of disease man-
agement. By analyzing sentiment trends in Arabic sentiment
text, authorities can prioritize resources, allocate personnel,
and tailor intervention strategies based on the sentiment pat-
terns observed. This optimizes efforts and ensures efficient
utilization of limited resources.
3 Methodology
This section describes the proposed method of DNN that
is used for classification and ML algorithms that are used
for comparing with the proposed approach. Figure1 shows
the framework started from data collection and preprocess-
ing, which are removing duplicated values, replacing each
emoji with text, removing Arabic stop words, removing non-
Arabic words, removing diacritics and numbers and punc-
tuation, removing hashtags and links, and reducing words to
their roots, and the end of preprocessing is extract features
from text using Count Vectorizer and TfidfTransformer
and SMOTE. Final classical ML algorithms are applied
to the data, which are Light Gradient Boosting, AdaBoost,
XGBoost, and SVM. Also, differential evolution optimiza-
tion is used with the ML algorithm, and then, the proposed
DNN approach is applied.
Table 1 Summary of the previous work
References Technique Dataset Best result
Atoum and Nouman (2019) Naive Bayes and SVM 1000 tweets collected from Twitter Accuracy of 82.1%
Tan etal. (2023) Multi-Layer Perceptron (MLP)
machine learning model
Emotional data Accuracy of 88.7%
Rukhsar etal. (2023) LSTM and SVM 90,000 tweets relevant to the
COVID-19
Accuracy of 90%
Heikal etal. (2018) LSTM and CNN Arabic sentiment tweets dataset
(ASTD)
F1-score: 64.46%
Sayed etal. (2020) NB, KNN, MLP, RF, SVM, DT,
GB, RC, and LR
6318 reviews written in various
forms of the Arabic language
95.21%
Alayba etal. (2018) LSTM and CNN Arabic health-care services dataset,
Twitter dataset, and Arabic senti-
ment tweets dataset
First database: 94.24%, second
database: 95.68%, and third
database: 92%
Oussous etal. (2020) CNN, LSTM, SVM, and Naïve
Bayes
Moroccan Sentiment Analysis
Corpus (MSAC)
96%
Hadwan etal. (2022) Naïve Bayes, DT, SVM, and KNN The dataset was collected contains
7759 reviews
78.46%
Baker etal. (2020) KNN, Naïve Bayes, SVM, and DT 6300 tweets about influenza 89.06%
Gamal etal. (2019) SVM, Naïve Bayes, Ridge Regres-
sion, and AdaBoost
151,000 tweets collected about vari-
ous Arabic dialects
99.9%
Mohammed and Kora (2019) LSTM, CNN, and RCNN 40,000 tweets were collected 81.31%
Aloqaily etal. (2020) LMT, KNN, SVM, Vote, and Sim-
ple Logistic
Arabic tweets dataset Acc = 85.55%, F1 = 92%, and
AUC = 86%
Hnaif etal. (2021) SVM, Naïve Bayes, KNN, Random
Forest, and Decision Tree
6138 tweets and posts collected
from Facebook and Twitter
F1-score: 83%
Al-Tamimi etal. (2017) Bernoulli NB, SVM-RBF, and KNN 8053 comments from Youtube 88.8%
Abu-Farha and Magdy (2020) BiLSTM ArSarcasm (10,547 tweets) 46%
Alwakid etal. (2017) SVM and Naïve Bayes 4000 tweets were collected 73%
Alayba etal. (2017) Logistic regression, Stochastic Gra-
dient Descent, CNN, SVM, and
Naïve Bayes
2026 tweets 90.14%

Social Network Analysis and Mining (2024) 14:30 Page 5 of 18 30
3.1 Data collection
The data that are used in this work are collected using
Twitter API and Tweepy library. Tweepy, a Python module
for accessing the Twitter API, provides developers with
access to Twitter content, such as tweets, retweets, and
timestamps. The focal point in choosing Monkeypox was
because of is an epidemic that spread quickly and takes a lot
of interest, especially in the Arab world, with the selection of
Arab tweets related to these polarizing incidents for analysis.
The data collection phase extended for eight months from
28/7/2022 to 4/4/2023, when 4763 tweets were collected and
classified by language specialists as positive, negative, and
neutral based on the meaning of the sentence, where 1102
tweets were neutral, 944 tweets were positive, and 2717
were negative. The data contained Modern Standard Arabic
(MSA) and Arabic dialects but the most was Arabic dialects.
Some samples of data with translation are shown in Table2.
3.2 Data preprocessing
In the realm of natural language processing (NLP), data pre-
processing encompasses a series of steps taken to cleanse,
convert, and ready raw text data for analysis and modeling
purposes. The ensuing information outlines various preproc-
essing tasks specific to NLP.
3.2.1 Eliminating duplicate values
This stage entails identifying and eliminating instances of
redundant data. Duplicates may arise when identical text
data appear multiple times within the dataset. By removing
duplicates, we ensure that each unique instance is repre-
sented only once, thereby averting data redundancies. This
step was performed twice, both prior to and after preprocess-
ing, taking into account the presence of duplicated tweets
with different emojis or hashtags. After removing these dis-
crepancies, the tweets became identical.
Fig. 1 The proposed methodol-
ogy of work
Table 2 Some instances for tweets and their translate with their label
Tweet text Translate of tweet text Label
          
           
         
Investors are racing to acquire shares of companies that may help eradi-
cate Monkeypox, which has become a source of concern, at a time when
health authorities around the world intensify their search to stop the
outbreak of the disease
Positive
          
          

New York declares a public emergency due to the high incidence of
"monkeypox," and the mayor of the city confirms that nearly a thousand
residents of the city are at risk of infection
Negative
          There is nothing but a well-known observation of Monkeypox, and those
who were before us knew it
Natural

Social Network Analysis and Mining (2024) 14:30 30 Page 6 of 18
3.2.2 Removing Arabic stop words
In NLP, it is customary to eliminate stop words to reduce
data dimensionality and prioritize more meaningful words.
Arabic stop words are specific to the Arabic language and
comprise articles, prepositions, pronouns, and similar terms.
By eliminating these stop words, we focus on the most per-
tinent content words.
3.2.3 Removing diacritics, numbers, punctuation,
hashtags, andlinks
Diacritics encompass marks or symbols added to charac-
ters to indicate pronunciation or other linguistic features.
Removing diacritics involves stripping these markings
from the text. Removing numbers and punctuation serves
to simplify the text and remove potentially irrelevant or dis-
tracting elements. Hashtags and links frequently appear in
social media or web-based data, and their removal facilitates
concentration on the textual content rather than metadata or
web-related information.
3.2.4 Emojis extracting
Replacing each emoji with text, emojis are graphical sym-
bols used to express emotions or convey ideas in written
communication to enhance the interpretability of NLP algo-
rithms. In this work, we extracted all the emojis from the
existing texts, then defined each emoji with the Arabic word
that represents it, and then, all the emojis were replaced
according to the definition that was written.
Algorithm1 Pseudo-code of preprocessing steps
Input: Dataset
Output: Cleaned data
1. Begin
2. data <- Remove Duplicates(data)
3. data <- Replace Emoji with Text(data)
4. data <- ExtractEmoji(data)
5. constants <- Load EmojiConstants()
6. data <- Replace Emoticons with Emoji(data)
7. data <- Remove Arabic Stopwords(data)
8. data <- Remove Non-Arabic Words(data)
9. data <- Remove Diacritics(data)
10. data <- Remove Numbers(data)
11. data <- Remove Hashtags(data)
12. data <- Remove Links(data)
13. data <- Remove Punctuations(data)
14. data <- ReduceWordsToRoots(data)
15. data <- Remove Duplicates(data)
16. End
3.2.5 Stemming words
These techniques aim to convert words back to their base or
root form, thereby reducing variations of the same word to
a common representation. For instance, the words “”
(writing), “” (written), and “” (writes) would be
reduced to the root form “” (write).
3.2.6 Extracting features fromtext using CountVectorizer,
TfidfTransformer, andSMOTE
Extracting features from text entails converting textual
data into numeric representations that can be processed
by machine learning algorithms. CountVectorizer is a
method that transforms text into an array, indicating the
frequency of words in the text. TfidfTransformer calculates
term frequency-inverse document frequency (TF-IDF) val-
ues, reflecting the significance of words in a document
relative to the entire corpus. SMOTE (Synthetic Minority
Oversampling Technique) (Chawla etal. 2002) (Kovács
etal. 2020) is a technique employed to address class
imbalance in datasets by oversampling the minority class,
thus enhancing the performance of classification models.
3.3 Machine learning algorithms
In this research, we have created, built, and evaluated mul-
tiple models using various machine learning techniques
which are LGBM, AdaBoost, XGBoost, and SVM. We
divided 70% of the datasets for training, 15% for valida-
tion, and 15% for testing purposes. The default hyper-
parameter settings in sklearn were used for all classical
machine learning algorithms. Then, differential evolution
optimization algorithms are used to raise the accuracy by
adjusting the hyperparameters in each machine learning
algorithm. The next section provides a discussion of the
implemented techniques.
3.3.1 Light gradient boosting machine (LGBM)
LightGBM is also a learning algorithm that enhances gradi-
ent and is similar to XGBoost in terms of efficiency, speed,
and performance. This model is distinguished by its work
on reducing data in training through the use of a technique
called Gradient-based One-Side Sampling (GOSS), where
the leaf growth method is also used. Unlike other enhanced
ranking algorithms increase its performance and efficiency
in classification (Ke etal. 2017).
3.3.2 Adaptive boosting (AdaBoost)
AdaBoost is a classification algorithm based on gradient
reinforcement and like XGBoost. Its working principle is

Social Network Analysis and Mining (2024) 14:30 Page 7 of 18 30
based on choosing a strong classifier from a group of weak
classifiers. In each iteration, the weights of the erroneous
classifications are adjusted to be corrected in the next itera-
tion, which contributes to enhancing efficiency and obtain-
ing high accuracy (Margineantu and Dietterich 1997).
3.3.3 Support vector machine (SVM)
SVM is a learning algorithm that works on classification and
regression, where it works to separate features by finding the
best hyperplane and also works to increase generalization by
widening the margin that separates the decision limit from
the nearest data point, and SVM can be suitable for high-
dimensional data (Jakkula 2011).
3.3.4 Extreme gradient boosting (XGBoost)
XGBoost is a machine learning algorithm for scaling up that
is characterized by speed and performance. It relies in its
work on building weak decision trees to create strong pre-
dictive decisions. Some decision trees are added iteratively,
which contributes to improving the loss function, which
helps prevent overfitting (Chen etal. 2018).
3.3.5 Hyperparameter tuning using differential evolution
The differential evolution (DE) algorithm (Derviş and
Selçuk 2004) is a heuristic method that offers three
important advantages: It can find the true global minimum
regardless of the values of the initial parameters, it exhibits
fast convergence, and it only requires a few control
parameters. DE is a population-based algorithm that is
similar to genetic algorithms and uses common factors such
as crossover, mutation, and selection. The main difference
between genetic algorithms and DE lies in their approach
to building better solutions. While genetic algorithms rely
heavily on the intersection, DE focuses on the process of
mutation.
The DE mutation process depends on the differences
between pairs of solutions taken at random in the population.
This process acts as a research mechanism, while the selec-
tion process directs the research toward the most promising
areas in the field of research.
In DE, the optimization task is represented using D
parameters as a D-dimensional vector. The algorithm begins
by randomly generating a set of NP solvent vectors. These
vectors are then iteratively optimized by applying mutation,
crossover, and selection factors. This iterative process aims
to find the best solution for the optimization problem in the
specified search space.
For each target vector
xi
,
G
a mutant vector is produced by:
where
i,r1,r2,r3
∈ {1, 2,…,NP} are randomly chosen and
must be different from each other. In Eq.(1), F is the scal-
ing factor which has an effect on the difference vector
(xr2, G − xr3, G), and K is the combination factor. The DE
algorithm also uses non-uniform crossover which gives pri-
ority to taking child vector parameters from one parent more
frequently than the other. Using components from existing
community members were to build beta vectors. The recom-
bination factor (intersection) efficiently switches the infor-
mation about successful combinations making it easier to
search for better solution spaces. The parent vector is mixed
with the mutated vector to produce a trial vector
uj,i,G+1
where
j=1, 2, ..., D; rj ∈[0, 1]
is the random number;
CR
is crossover constant
∈[0, 1]
; and
rni∈ (1,2, ..., D)
is the
randomly chosen index.
All solutions in the community have an equal oppor-
tunity to choose parents without relying on their fitness
value. The child produced after the mutation and crosso-
vers is evaluated. After that, the performance of the child
and its parent vectors is compared, and the best is chosen.
If the parent is still better, it is kept among the population.
Algorithm2 shows steps of DE optimization.
Algorithm2 Pseudo-code of DE optimization
1. Initialization
2. Repeat
3. Mutation
4. Crossover
5. Evaluation
6. Selection
7. Until (termination criteria are met)
(1)
vi,G+1
=x
i,G
+K⋅
(
x
r1,G
−x
i,G)
+F⋅
(
x
r2,G
−x
r3,G)
(2)
u
j,i,G+1=
{
vji,G+1if
(
rndj≤CR
)
or j=rn
i
q
ji,G
if(rnd >CR)and j≠rn
i
Table 3 Hyperparameters tuning for all models
ML model Best hyperparameters
SVM C: 7.59
Kernel: rbf
Gamma: 0.68
AdaBoost Learning rate: 0.319 n_estimators: 300
XGBoost Learning rate: 0.433 max_depth: 9
LGBM Learning rate: 0.542 max_depth: 8

Social Network Analysis and Mining (2024) 14:30 30 Page 8 of 18
Table3 shows hyperparameter tuning for different
machine learning models (SVM, AdaBoost, XGBoost,
and LGBM) and identified the best hyperparameters for
each model.
3.4 Proposed DNN approach
An artificial neural network (ANN) approach was used as a
computational model because it is influenced by the proper-
ties of biological neural networks to incorporate intelligence
into our proposed method. Feed-Forward Neural Network
(FFN), a type of ANN, is represented as a graph directed
to pass various system information along the edges from
one node to another without forming a cycle (Ullah and
Mahmoud 2022). We adopt a multi-layer model (MLP),
which is a kind of FFN in the proposed model consisting
of one input layer, three hidden layers, and an output layer
containing three labels; each layer contains many neurons
Fig. 2 DNN architecture
Fig. 3 Flowchart for ReLU, Leaky ReLU, and ReLU6

Social Network Analysis and Mining (2024) 14:30 Page 9 of 18 30
or units in mathematical notation. Experiments choose the
number of hidden layers. The information is transferred
from one layer to the other in a forward direction, with
neurons in each layer fully connected, as shown in figure.
It starts from 7008, which is the input and the count of fea-
tures, then there is output of it which are 512, and we call
that the input is fully connected layer, and there are three
hidden fully connected layers, which are (512, 256), (256,
128), and (128, 32) followed by final output fully connected
layer which is (32, 3) noting that 3 is the count of classes.
Also, after every fully connected layer, there is an activa-
tion function. Figure2 shows the general architecture of
DNN methods.
MLP is defined mathematically as
Rm×Rn
where m is the
size of inputs
x=x1,x2,x3,….., xm
, and n is the outputs size
hi(x)
and defined mimetically as follows:
Which f is ReLU, Leaky ReLU, and relu6 activation func-
tions, which are used as three experiments to reduce the
state of vanishing and error gradient issue also to parsing
between them. Each function is used after every layer. This
way of stacking hidden layers is typically called deep neural
networks (DNNs) as shown in Fig.2. The activation func-
tions are defined mathematically as the equation below and
as shown in Fig.3.
SReLU (Scaled Exponential Linear Unit): The SReLU,
or Scaled Exponential Linear Unit, is an activation function
that introduces a piecewise linear curve with two parameters:
alpha and beta. The mathematical expression for SReLU is
as follows:
The SReLU is like the ReLU (Rectified Linear Unit) in
that it only activates for positive input values, setting nega-
tive values to zero. However, the SReLU has two additional
parameters (alpha and beta) that allow it to have different
(3)
hi
(x)=f
(
w
t
i
x+bi
)
(4)
ReLU = max (0, x)
slopes for positive input values, enabling more flexibility
in modeling complex data distributions (Gustineli 2022).
Leaky ReLU: The Leaky ReLU is a variant of the tradi-
tional ReLU that addresses a potential issue called “dying
ReLU.” In ReLU, neurons with negative inputs are assigned
an output of zero, effectively deactivating them. If a neuron
is deactivated for all inputs during training, it may become
stuck in that state and not update its weights further, leading
to the neuron “dying” and not contributing to the network’s
learning process (Gustineli 2022; Apicella etal. 2021). The
mathematical expression for Leaky ReLU is as follows:
In the Leaky ReLU, instead of setting negative val-
ues to zero, it introduces a small slope (0.1 in this case)
for negative inputs. This small slope allows the neuron to
carry a small gradient even for negative inputs, preventing
the “dying ReLU” problem and promoting better learning
in deeper neural networks. Figure3 shows a flowchart for
ReLU, Leaky ReLU, and ReLU 6.
ReLU6: ReLU6 is another variant of the ReLU activation
function that bounds the output to a maximum value, usu-
ally 6 (Kim etal. 2021). The mathematical expression for
ReLU6 is as follows:
It behaves like a regular ReLU for positive inputs, set-
ting them to the input value. However, it clips the output to
6 if the input value is greater than 6. This bounded behavior
can be helpful in preventing extremely large activations
that might cause numerical instability or other issues in
certain neural network architectures. The provided Algo-
rithm3 represents a simple neural network model with a
Leaky ReLU activation function. The neural network is
designed for a multi-class classification problem with three
output classes. Below is a brief explanation of theneural
net model.
(5)
Leaky ReLU =max(0.1x,x)
(6)
ReLU6 =min(max (0, x),6
)

Social Network Analysis and Mining (2024) 14:30 30 Page 10 of 18
Algorithm3 Pseudo-code of neural net model
Input:
X_train (input data), shape (n_samples, n_features)
Output:A deep neural network model: NeuralNet
1. Begin
//Data Collection: Data were gathered using Twitter API and Tweepy library which are 4763 tweets divided
to 1102 tweets were neutral, 944 tweets were positive and 2717 were negative)
2. Data <- Preprocessing
3. function NeuralNet():
4. fc1 <- Linear layer with input size matching the number of features and output size 512
5. fc2 <- Linear layer with input size 512 and output size 256
6. fc3 <- Linear layer with input size 256 and output size 128
7. fc4 <- Linear layer with input size 128 and output size 32
8. fc5 <- Linear layer with input size 32 and output size 3
//(3 output classes)
9. forward function:
10.Input: x
11.Output: x
12.x = Leakyrelu(fc1(x))
13.x = Leakyrelu (fc2(x))
14.x = Leakyrelu (fc3(x))
15.x = Leakyrelu (fc4(x))
16.x = fc5(x)
17.x = Leakyrelu (x)
18.return x
19.return forward
20.function train(X_train):
21.model = NeuralNet()
22.optimizer = Adam(model.parameters())
23.criterion = CrossEntropyLoss()
24.for epoch in range(num_epochs):
25.optimizer.zero_grad()
26.outputs = model.forward(X_train)
27.loss = criterion(outputs, y_true)
28.loss.backward()
29.optimizer.step()
30.function test(X_test):
31.model = NeuralNet()
32.outputs = model.forward(X_test)
33. predicted_labels = argmax(outputs, axis=1)
34.return predicted_labels
35.
End
3.5 Performance models
Performance evaluation of machine learning models is a cru-
cial step in assessing their effectiveness in solving specific
tasks. In this research, the model’s evaluation is based on
several performance metrics accuracy, initiative, recall, and
F1-score with a preference for F1-score and accuracy.
Accuracy: It is the most used, and probably the first choice,
for evaluating an algorithm’s performance on classification
problems. It is defined as the ratio of true classifications to all
true or false classifications (Eusebi 2013).
Precision: It simply refers to “the number of specific
relevant data items,” as the name implies. Or how many
positive notes the algorithm predicted were positive (Juba
and Le 2019). Precision is equal to the number of true
positives divided by the total number of true positives and
false positives:
(7)
Accuracy
=
TP +TN
TP +FP +FN +TN
(8)
Precision
=
TP
TP +FP

Social Network Analysis and Mining (2024) 14:30 Page 11 of 18 30
Recall: How many positive observations—really cor-
rect ones—have the algorithm predicted? Recall equals the
number of true positives divided by the total number of true
positives and false negatives (Otten etal. 2005).
F1-score: This metric, often known as an f-score, evalu-
ates the accuracy and callback of an algorithm to evaluate its
performance. It is theoretically represented as the harmonic
mean of recall and accuracy (Fourure etal. 2021) which
represents the best evaluation in some classifications.
(9)
Recall
=
TP
TP +FN
4 Results anddiscussion
In this study, we will discuss a comprehensive exploration of
the performance characteristics of various machine learning
algorithms also using different evolution optimization algo-
rithms with compressing with proposed DNN model. The
performance of each model is evaluated based on accuracy,
(10)
F1
−score =2×
precision ×recall
precision +recall
Table 4 Performance of DNN
and ML algorithms with DE
and without it
Type of algorithm Performance Avg precision Avg recall Avg F1-score Accuracy
ML algorithm without DE SVM 0.86 0.86 0.86 0.86
LGBM 0.85 0.85 0.85 0.86
AdaBoost 0.74 0.73 0.73 0.73
XGBoost 0.85 0.85 0.85 0.85
ML algorithm with DE SVM 0.90 0.90 0.90 0.90
LGBM 0.85 0.85 0.85 0.86
AdaBoost 0.80 0.79 0.79 0.79
XGBoost 0.87 0.86 0.86 0.87
Proposed DNN DNN + ReLU 0.9109 0.9091 0.9105 0.9125
DNN + Leaky ReLU 0.9169 0.9165 0.9167 0.9182
DNN + ReLU6 0.9109 0.9104 0.9106 0.9125
Fig. 4 Performance comparison
of ML algorithm

Social Network Analysis and Mining (2024) 14:30 30 Page 12 of 18
precision, recall, and F1-score, the result of classical ML
algorithm shown in table. Our goal is to shed light on how
these algorithms perform in different contexts and to under-
stand the factors that contribute to their effectiveness. To
achieve this, we categorize our analysis into three main sec-
tions: machine learning (ML) algorithms without optimiza-
tion, ML algorithms with optimization, and Deep Neural
Network (DNN) algorithms as shown in Table4. Each of
these categories represents a distinct approach to handling
predictive tasks, and our aim is to delve into the intricacies
of their performances.
The initial category centers on machine learning algo-
rithms that operate without any optimization. This phase
of our study is designed to assess the inherent capabilities
of these algorithms and identify those that exhibit the most
promising metrics.
Four outstanding algorithms Support Vector Machine
(SVM), LightGBM (LGBM), AdaBoost, and XGBoost are
subjected to a rigorous evaluation process. SVM, a widely
employed classification algorithm, consistently demon-
strates performance across all metrics, yielding an average
precision, recall, F1-score, and accuracy of 0.86 as shown
in Fig.4. Although it may not reach the pinnacle of optimi-
zation, SVM stands out for its reliability and ability to pro-
duce well-balanced outcomes. Similarly, LGBM proves to
be reliable, boasting an average precision, recall, F1-score,
and accuracy of 0.85. This highlights LGBM’s capability
to generate dependable predictions even without extensive
optimization, positioning it as a versatile choice across
diverse applications.
Contrastingly, AdaBoost’s performance metrics average
precision, recall, F1-score, and accuracy register at 0.74,
slightly trailing the top-performing algorithms. However,
the potential of AdaBoost in specific scenarios is not to
be underestimated. Its specialized applicability makes
it a viable consideration for tasks where its strengths can
shine. The XGBoost algorithm, recognized for its ensemble
approach, consistently performs with an average precision,
recall, F1-score, and accuracy of 0.85. This underscores its
suitability for tasks that demand well-balanced and reliable
predictions.
Transitioning to the second category, we delve into the
realm of machine learning algorithms with optimization.
This phase shifts our focus toward understanding the impact
of optimization on algorithmic performance. In Fig.5, we
reevaluate the four previously mentioned algorithms SVM,
LGBM, AdaBoost, and XGBoost after subjecting them to
optimization, aiming to uncover potential improvements in
their predictive capabilities. Post-optimization, SVM experi-
ences a substantial performance boost, achieving an average
precision, recall, F1-score, and accuracy of 0.9. This notable
enhancement highlights SVM’s adaptability for refined pre-
dictions and further solidifies its standing as a robust con-
tender. LGBM, even after optimization, maintains its perfor-
mance integrity, with an average precision, recall, F1-score,
and accuracy of 0.85. This consistency underscores LGBM’s
inherent reliability and stability.
Optimization proves to be a crucial factor in elevating
AdaBoost’s metrics, resulting in an average precision, recall,
F1-score, and accuracy of 0.8. This positions AdaBoost as
Fig. 5 Performance comparison
of ML algorithm with optimiza-
tion

Social Network Analysis and Mining (2024) 14:30 Page 13 of 18 30
a valuable choice, particularly in scenarios where optimiza-
tion plays a pivotal role. Similarly, the XGBoost algorithm,
following optimization, exhibits heightened results, boasting
an average precision, recall, F1-score, and accuracy of 0.87.
This advancement further solidifies XGBoost’s adaptability
for precise predictions, reaffirming its place in the landscape
of machine learning algorithms.
In the final phase of our analysis as shown in Fig.6,
we introduce the realm of Deep Neural Network (DNN)
algorithms. This part of the study focuses on assessing the
performance of DNN algorithms under various activation
Fig. 6 Performance comparison
of DL algorithm with optimiza-
tion
Fig. 7 Receiver operating
characteristic (ROC) for Leaky
ReLU

Social Network Analysis and Mining (2024) 14:30 30 Page 14 of 18
functions to understand their impact on predictive capa-
bilities. The first DNN variant, DNN + ReLU, demon-
strates commendable performance, achieving an average
precision, recall, F1-score, and accuracy of 0.9109. This
consistent and high-scoring performance underscores its
competence in predictive tasks. Surpassing its predeces-
sor, DNN + Leaky ReLU records an average precision,
recall, F1-score, and accuracy of 0.9169. This improve-
ment highlights the efficacy of the Leaky ReLU activa-
tion function in enhancing overall performance. Simi-
larly, DNN + ReLU6 attains an average precision, recall,
F1-score, and accuracy of 0.9109, mirroring the stability
and reliability of the ReLU6 activation function.
For the proposed model, a higher accuracy was obtained
than the rest of the models. During the three experiments,
the highest accuracy was achieved by using the model
with Leky ReLU how much appears in the table, followed
by the use of relu6. In the end, ReLU with the proposed
model achieved less accuracy, and despite that it outper-
formed the rest of the models machine learning. The rea-
son why Leaky ReLU is superior to ReLU and relu6 is to
avoid the problem of dying ReLU. Instead of converting
negative values into zeros and causing inactivity of some
neurons, they turn into very small negative values, and
because of that, a slowdown in the calculation process
can occur.
Fig. 8 Confusion matrix for DNN with activation function that used

Social Network Analysis and Mining (2024) 14:30 Page 15 of 18 30
By examining the box plot in the figure, we can gain
insights into the performance differences and distributions
between different machine learning models for each met-
ric. We can see that the proposed DNN model has the best
results. Also, the figure displays the confusion matrix of
the DNN model with the three activation functions. When
analyzing the confusion matrix, we observed a large number
of true positive for certain categories, indicating that the
model correctly predicted these cases. This makes the model
effective for detecting positive cases. An AUC value of 0.97
indicates that the model is highly effective in differentiating
between positive and negative instances as shown in Fig.7.
It suggests that the model has a high true-positive rate (sen-
sitivity) and a low false-positive rate (1-specificity).
This could be attributed to the model’s difficulty in recog-
nizing subtle differences in text that indicates a more neutral
or negative sentiment. As shown in Fig.8, falsely classi-
fied as negative sentiments were 41 instances, whereas they
should have been categorized as neutral or positive. The mis-
classifications may be a result of complex language usage,
sarcasm, or nuanced expressions that the model struggled
to comprehend accurately. As the field of NLP continues to
advance, minimizing false positives will play a significant
role in harnessing the true potential of sentiment analysis
in diverse applications, from market research to customer
support and beyond.
When comparing our study to previous research (Ke,
etal. 2017; Margineantu and Dietterich 1997; Jakkula
2011; Chen etal. 2018; Derviş and Selçuk 2004; Ullah and
Mahmoud 2022; Gustineli 2022; Apicella etal. 2021; Kim
etal. 2021; Eusebi 2013; Juba and Le 2019; Otten etal.
2005; Fourure etal. 2021), certain distinctive features come
to light. Our study stands out due to its specific emphasis on
Arabic COVID-19 tweets and its utilization of a Deep Neural
Network (DNN) approach as shown in Table5. Our research
bears several noteworthy attributes that differentiate it from
the pre-existing body of work, making substantial contribu-
tions to the domain of disease-related social media analysis.
Particularly, the distinct focus on Arabic COVID-19 tweets
distinguishes our study. Whereas preceding research encom-
passed a variety of diseases and languages, our concentrated
investigation into a specific ailment within a particular lin-
guistic community offers illuminating perspectives on how
COVID-19 is perceived and discussed among Arabic speak-
ers. This unique focus unveils cultural and linguistic nuances
that can offer deeper insights into pandemic perceptions.
The superiority of some research to the results obtained
is due to the lack of data collected and limited to one field,
which is Monkeypox, in contrast with the previous research
that worked on datasets in a broader field. In addition to the
great similarity in positive, negative, and natural tweets, it
affects the classification process. Given that the data col-
lected is real and that the tweets were written in colloquial
Arabic from more than one Arab country, and thus, there
is diversity in Arabic dialects, working on it is sometimes
Table 5 Performance of related work on Arabic and English sentiment analysis datasets
References Methodology Appli-
cation
language
Accuracy (%)
Abdurrahim etal. (2022) Gaussian Naïve Bayes, Bernoulli Naïve Bayes, and Naïve Bayes Multino-
mial
English 80
Mohbey etal. (2022) CNN and LSTM English 94
Bengesi etal. (2023) Random Forest, Logistic Regression, Multi-layer Perceptron (MLP), SVM,
KNN, Naïve Bayes, and XGBoost
English 93.48
Oscar etal. (2017) ML classifiers English 95.15
Chintalapudi etal. (2021) Logistic regression, SVM, and LSTM English 87.33
Dangi etal. (2022) Decision Tree, SVM, Logistic Regression, Random Forest, and Multino-
mial NB
English 98
Iparraguirre-Villanueva etal. (2023) CNN and LSTM English 88
Musleh etal. (2022) SVM, Random Forest (RF), Logistic Regression (LR), KNN, AdaBoost,
and Naïve Bayes (NB)
Arabic 82.39
Al-Musallam and Al-Abdullatif (2022) SVM, KNN, Logistic Regression (LR), and Naïve Bayes Arabic 82
Baker etal. (2020) Naive Bayes, SVM, KNN, and Decision Trees Arabic 89.06
Aljameel etal. (2021) SVM, KNN, and Naïve Bayes Arabic 85
Waheeb etal. (2020) Random Forest, SVM, CNN (LSTM), RNN (LSTM), and Naïve Bayes Arabic 89
Boulesnane etal. (2022) SVM, BiLTSM, GRU, and LSTM Arabic 89.8
Alabid and Katheeth (2021) Support Vector Machine (SVM) and Naive Bayes Arabic 81%
Our research DNN Arabic 92

Social Network Analysis and Mining (2024) 14:30 30 Page 16 of 18
difficult, and the extracted features are considered more
complex.
Due to the previous work, most researchers used ML in
their work and did not expand their methodology. Therefore,
the proposed approach is considered distinct from the rest, as
it contains an effective preprocessing to prepare the data for
training, in addition to the use of DNN with three different
classification methods, and the inclusion of DE to help in
adjusting the parameters.
Furthermore, the adoption of a Deep Neural Network
(DNN) approach sets our research on the vanguard of con-
temporary methods for deciphering social media data. DNNs
have proven exceptionally adept at uncovering intricate pat-
terns and intricate connections within data, making them
particularly apt for capturing the subtleties of language and
sentiment prevalent in social media dialogs. The adoption
of this modern approach elevates our research by enhanc-
ing the accuracy and depth of our analysis. Consequently,
while our attained accuracy score of 92% resides within the
range observed in prior studies, it remains an achievement
worthy of commendation, especially given the intricate
nature of deciphering social media data. This attainment
underscores the fact that our chosen DNN approach is on
par with or potentially surpasses the accuracy exhibited by
other methodologies employed in earlier investigations. Such
a high accuracy score is pivotal in establishing the credibility
and validity of our findings. Above all, our research carries
significant implications for the expansive realm of disease-
related social media analysis. With its laser focus on Arabic
COVID-19 tweets, our study delves into a previously unex-
plored facet of disease dialogues. This, in turn, augments our
comprehension of public perceptions and sentiments, while
concurrently delivering invaluable insights for shaping public
health interventions and tailoring policies to resonate with
the Arabic-speaking population. The distinctive aspects of
our research its focal point on Arabic COVID-19 tweets,
incorporation of a DNN methodology, competitive accuracy
score, and the absence of explicitly stated limitations jointly
position it as a valuable addition to the field of disease-related
social media analysis. The strengths of our study not only
underscore its capacity to deepen our comprehension of dis-
ease discussions within the Arabic-speaking community but
also to provide actionable insights for health-care profession-
als and researchers.
5 Conclusion
This study focused on the application of sentiment analysis
in analyzing Monkeypox using Arabic sentiment text, with
an emphasis on hyperparameter optimization for machine
learning and deep learning algorithms. The deep learning
model-based Leaky ReLU showcasing the significance of
leveraging complex representations for NLP tasks with 92%.
The use of deep learning models outperformed traditional
machine learning algorithms, highlighting the importance
of leveraging complex representations for NLP tasks. By
gaining insights into public sentiments toward Monkeypox,
health-care authorities and policymakers can make more
informed decisions to address public concerns effectively.
This research contributes to advancing sentiment analysis
techniques in Arabic text and sheds light on the potential
applications of sentiment analysis in the context of public
health issues. As the field of sentiment analysis continues
to evolve, these models can be further refined to yield even
more accurate and efficient results, ultimately aiding in
decision-making processes, and enhancing public health
outcomes.
Limitation: Research limitations center on the difficulty
of collecting data, as it takes a long time, in addition to its
scarcity, given that the subject of Monkeypox is still new. In
addition, the preprocessing of the data took a great deal of time
and effort due to its difficulty in the Arabic language. In the
future, we seek to collect larger data in addition to including
more important topics to know people’s feelings about them.
Author contributions HG worked in software, resources, writing—
original draft, supervision, methodology, conceptualization, formal
analysis, and review and editing. REAM worked in supervision, meth-
odology, conceptualization, and writing—original draft. GS helped in
formal analysis, and writing—review and editing. AN helped in formal
analysis and writing—review and editing. SS helped in formal analysis
and writing—review and editing. KMON helped in formal analysis
and writing—review and editing. MA helped in formal analysis and
writing—review and editing. EAD helped in formal analysis and writ-
ing—review and editing. MG helped in formal analysis and writing—
review and editing. LA helped in formal analysis and writing—review
and editing. All authors read and approved the final paper.
Funding Not applicable.
Data availability Data are available from the authors upon reasonable
request.
Declarations
Conflict of interest The authors declare that there is no conflict of in-
terest regarding the publication of this paper.
Ethical approval This article does not contain any studies with human
participants or animals performed by any of the authors.
Informed consent Informed consent was obtained from all individual
participants included in the study.
References
Abdurrahim A, Syafa’ah L, Lestandy M (2022) Sentiment analysis of
Covid-19 vaccine tweets utilizing Naïve Bayes. AIP Conference
Proceedings, vol. 2453. https:// doi. org/ 10. 1063/5. 00946 07

Social Network Analysis and Mining (2024) 14:30 Page 17 of 18 30
Abu-Farha I, Magdy W (2020) From Arabic sentiment analysis to
sarcasm detection: the ArSarcasm dataset, Aclweb.Org, Euro-
pean L, pp 32–39
Alabid NN, Katheeth ZD (2021) Sentiment analysis of twitter posts
related to the covid-19 vaccines. Indones J Electr Eng Comput
Sci 24(3):1727–1734. https:// doi. org/ 10. 11591/ ijeecs. v24. i3.
pp1727- 1734
Alayba AM, Palade V, England M, Iqbal R (2017) Arabic language
sentiment analysis on health services, pp 114–118. https:// doi.
org/ 10. 1109/ asar. 2017. 80677 71
Alayba AM, Palade V, England M, Iqbal R (2018) A combined CNN
and LSTM model for Arabic sentiment analysis. In: Lect. Notes
Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect.
Notes Bioinformatics), vol. 11015 LNCS, pp 179–191. https://
doi. org/ 10. 1007/ 978-3- 319- 99740-7_ 12
Aljameel SS etal (2021) A sentiment analysis approach to predict
an individual’s awareness of the precautionary procedures to
prevent covid-19 outbreaks in Saudi Arabia. Int J Environ Res
Public Health 18(1):1–12. https:// doi. org/ 10. 3390/ ijerp h1801
0218
Al-Musallam N, Al-Abdullatif M (2022) Depression detection through
identifying depressive Arabic tweets from Saudi Arabia: machine
learning approach. In: Proceedings 2022 5th National Conference
Saudi Computer Colleges, NCCC 2022, pp 11–18. https:// doi. org/
10. 1109/ NCCC5 7165. 2022. 10067 346
Aloqaily A, Al-Hassan M, Salah K, Elshqeirat B, Almashagbah M
(2020) Sentiment analysis for Arabic tweets datasets: Lexicon-
based and machine learning approaches. J Theor Appl Inf Technol
98(4):612–623
Al-Tamimi AK, Shatnawi A, Bani-Issa E (2017) Arabic sentiment
analysis of YouTube comments. In: 2017 IEEE Jordan confer-
ence on applied electrical engineering and computing technologie-
sAEECT, pp 1–6. https:// doi. org/ 10. 1109/ AEECT . 2017. 82577 66
Alwakid G, Osman T, Hughes-Roberts T (2017) Challenges in sen-
timent analysis for Arabic social networks. Proc Comput Sci
117:89–100. https:// doi. org/ 10. 1016/j. procs. 2017. 10. 097
Apicella A, Donnarumma F, Isgrò F, Prevete R (2021) A survey on
modern trainable activation functions. Neural Netw 138(June):14–
32. https:// doi. org/ 10. 1016/j. neunet. 2021. 01. 026
Atoum JO, Nouman M (2019) Sentiment analysis of Arabic Jordanian
dialect tweets. Int J Adv Comput Sci Appl 10(2):256–262. https://
doi. org/ 10. 14569/ ijacsa. 2019. 01002 34
Baker QB, Shatnawi F, Rawashdeh S, Al-Smadi M, Jararweh Y (2020)
Detecting epidemic diseases using sentiment analysis of arabic
tweets. J Univers Comput Sci 26(1):50–70. https:// doi. org/ 10.
3897/ jucs. 2020. 004
Bengesi S, Oladunni T, Olusegun R, Audu H (2023) A machine learn-
ing-sentiment analysis on Monkeypox outbreak: an extensive
dataset to show the polarity of public opinion from twitter tweets.
IEEE Access 11(February):11811–11826. https:// doi. org/ 10. 1109/
ACCESS. 2023. 32422 90
Boulesnane A, Meshoul S, Aouissi K (2022) Influenza-like illness
detection from Arabic Facebook posts based on sentiment analysis
and 1D convolutional neural network. Mathematics 10(21):1–22.
https:// doi. org/ 10. 3390/ math1 02140 89
Cambria E (2022) Sentic computing. In: Encyclopedia of big data, pp
821–827. Springer International Publishing, Cham
Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP (2002) SMOTE:
synthetic minority over-sampling technique. J Artif Intell Res
16:321–357
Chen T, He T, Benesty M (2018) XGBoost : eXtreme Gradient Boost-
ing. R Packag. version 0.71–2, pp 1–4
Chintalapudi N, Battineni G, Amenta F (2021) Sentimental analysis
of COVID-19 tweets using deep learning models. Infect Dis Rep
13(2):329–339. https:// doi. org/ 10. 3390/ IDR13 020032
Dangi D, Dixit DK, Bhagat A (2022) Sentiment analysis of COVID-
19 social media data through machine learning. Multimed
Tools Appl 81(29):42261–42283. https:// doi. org/ 10. 1007/
s11042- 022- 13492-w
Derviş K, Selçuk Ö (2004) A simple and global optimization algorithm
for engineering problems: differential evolution algorithm. Turk-
ish J Electr Eng Comput Sci 12(1):53–60
El-Beltagy SR, Ali A (2013) Open issues in the sentiment analysis of
Arabic social media: a case study. In: 2013 9th international con-
ference innovation information technology, IIT 2013, pp 215–220.
https:// doi. org/ 10. 1109/ Innov ations. 2013. 65444 21
Eusebi P (2013) Diagnostic accuracy measures. Cerebrovasc Dis
36(4):267–272. https:// doi. org/ 10. 1159/ 00035 3863
Fourure D, Javaid MU, Posocco N, Tihon S (2021) Anomaly detection:
how to artificially increase your F1-Score with a biased evalu-
ation protocol. In: Lecture notes computere science (including
Subseries lecture notes artifial intelligence lecture notes bioin-
formatics), vol. 12978 LNAI, pp 3–18. https:// doi. org/ 10. 1007/
978-3- 030- 86514-6_1
Gamal D, Alfonse M, El-Horbaty E-SM, Salem A-BM (2019) Twit-
ter benchmark dataset for Arabic sentiment analysis. Int J Mod
Educ Comput Sci 11(1):33–38. https:// doi. org/ 10. 5815/ ijmecs.
2019. 01. 04
Gustineli M (2022) A survey on recently proposed activation functions
for deep learning, pp 1–7. http:// arxiv. org/ abs/ 2204. 02921
Hadwan M, Al-Hagery MA, Al-Sarem M, Saeed F (2022) Arabic senti-
ment analysis of users’ opinions of governmental mobile applica-
tions. Comput Mater Contin 72(3):4675–4689. https:// doi. org/ 10.
32604/ cmc. 2022. 027311
Heikal M, Torki M, El-Makky N (2018) Sentiment analysis of Ara-
bic Tweets using deep learning. Proc Comput Sci 142:114–122.
https:// doi. org/ 10. 1016/j. procs. 2018. 10. 466
Hnaif AA, Kanan E, Kanan T (2021) Sentiment analysis for arabic
social media news polarity. Intell Autom Soft Comput 28(1):107–
119. https:// doi. org/ 10. 32604/ iasc. 2021. 015939
Huang H, Zavareh AA, Mustafa MB (2023) Sentiment analysis in
E-Commerce platforms: a review of current techniques and future
directions. IEEE Access 11(July):90367–90382. https:// doi. org/
10. 1109/ ACCESS. 2023. 33073 08
Iparraguirre-Villanueva O etal (2023) The public health contribution
of sentiment analysis of Monkeypox tweets to detect polarities
using the CNN-LSTM model. Vaccines 11(2):1–12. https:// doi.
org/ 10. 3390/ vacci nes11 020312
Jain R etal (2023) Explaining sentiment analysis results on social media
texts through visualization. Multimed Tools Appl 82(15):22613–
22629. https:// doi. org/ 10. 1007/ s11042- 023- 14432-y
Jakkula V (2011) Tutorial on Support Vector Machine (SVM). School
of EECS, Washington State University, pp 1–13
Juba B, Le HS (2019) Precision-Recall versus accuracy and the role
of large data sets. In: 33rd AAAI confernce artifial intelligence
AAAI 2019, 31st innovative applied artifial intelligence confer-
ence IAAI 2019 9th AAAI symposium education advance artifial
intelligence, EAAI 2019, pp 4039–4048. https:// doi. org/ 10. 1609/
aaai. v33i01. 33014 039
Ke G etal. (2017) LightGBM: a highly efficient gradient boosting
decision tree. Adv Neural Inf Process Syst, vol. 2017-Decem, pp
3147–3155
Kim H, Park J, Lee C, Kim JJ (2021) Improving accuracy of binary
neural networks using unbalanced activation distribution. In: Proc.
IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit., pp
7858–7867. https:// doi. org/ 10. 1109/ CVPR4 6437. 2021. 00777.
Kovács B, Tinya F, Németh C, Ódor P (2020) Unfolding the effects
of different forestry treatments on microclimate in oak forests:
results of a 4-yr experiment. Ecol Appl 30(2):321–357. https://
doi. org/ 10. 1002/ eap. 2043

Social Network Analysis and Mining (2024) 14:30 30 Page 18 of 18
Margineantu DD, Dietterich TG (1997) Pruning Adaptive Boosting
*** ICML-97 Final Draft ***"
Mohammed A, Kora R (2019) Deep learning approaches for Arabic
sentiment analysis. Soc Netw Anal Min 9(1):1–12. https:// doi. o rg/
10. 1007/ s13278- 019- 0596-4
Mohbey KK, Meena G, Kumar S, Lokesh K (2022) A CNN-LSTM-
based hybrid deep learning approach to detect sentiment polarities
on Monkeypox tweets, pp 1–11
Musleh DA etal (2022) Twitter arabic sentiment analysis to detect
depression using machine learning. Comput Mater Contin
71(2):3463–3477. https:// doi. org/ 10. 32604/ cmc. 2022. 022508
Oscar N, Fox PA, Croucher R, Wernick R, Keune J, Hooker K (2017)
Machine learning, sentiment analysis, and tweets: an examination
of Alzheimer’s disease stigma on Twitter. J Gerontol Ser B Psy-
chol Sci Soc Sci 72(5):742–751. https:// doi. org/ 10. 1093/ geronb/
gbx014
Otten JDM etal (2005) Effect of recall rate on earlier screen detection
of breast cancers based on the Dutch performance indicators. J
Natl Cancer Inst 97(10):748–754. https:// doi. org/ 10. 1093/ jnci/
dji131
Oussous A, Benjelloun FZ, Lahcen AA, Belfkih S (2020) ASA: a
framework for Arabic sentiment analysis. J Inf Sci 46(4):544–559.
https:// doi. org/ 10. 1177/ 01655 51519 849516
Rodríguez-Ibánez M, Casánez-Ventura A, Castejón-Mateos F, Cuenca-
Jiménez PM (2023) A review on sentiment analysis from social
media platforms. Expert Syst Appl. https:// doi. org/ 10. 1016/j. eswa.
2023. 119862
Rukhsar S etal (2023) Artificial intelligence based sentence level sen-
timent analysis of COVID-19. Comput Syst Sci Eng 47(1):791–
807. https:// doi. org/ 10. 32604/ csse. 2023. 038384
Salem F (2017) Social Media and the Internet of Things (The Arab
Social Media Report 2017), Arab Social Media Report Series
Sayed AA, Elgeldawi E, Zaki AM, Galal AR (2020) Sentiment analysis
for Arabic reviews using machine learning classification algo-
rithms. In: Proceedings 2020 international conference innovative
trends communications computere engineering, ITCE 2020, pp
56–63. https:// doi. org/ 10. 1109/ ITCE4 8509. 2020. 90478 22
Tan L, Tan OK, Sze CC, Bin Goh WW (2023) Emotional variance
analysis: a new sentiment analysis feature set for artificial intel-
ligence and machine learning applications. PLoS One 18(1):1–22.
https:// doi. org/ 10. 1371/ journ al. pone. 02742 99
Ullah I, Mahmoud QH (2022) An anomaly detection model for iot
networks based on flow and flag features using a feed-forward
neural network. In: Proceedings - IEEE consumer communica-
tions network conference CCNC, pp 363–368. https:// doi. org/ 10.
1109/ CCNC4 9033. 2022. 97005 97
Waheeb SA, Khan NA, Chen B, Shang X (2020) Machine learning
based sentiment text classification for evaluating treatment qual-
ity of discharge summary. Information. https:// doi. org/ 10. 3390/
INFO1 10502 81
Publisher's Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Springer Nature or its licensor (e.g. a society or other partner) holds
exclusive rights to this article under a publishing agreement with the
author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of
such publishing agreement and applicable law.
Authors and Aliations
HasanGharaibeh1· RabiaEmhamedAlMamlook2,3· GhassanSamara4· AhmadNasayreh1· SajaSmadi1·
KhalidM.O.Nahar1· MohammadAljaidi4· EssamAl‑Daoud4· MohammadGharaibeh5· LaithAbualigah6,7,8,9,10,11,12
* Laith Abualigah
Aligah.2020@gmail.com
Essam Al-Daoud
essamdz@zu.edu.jo
1 Department ofInformation Technology andComputer
Sciences, Yarmouk University, Irbid211633, Jordan
2 Department ofBusiness Administration, Trine University,
Indiana, USA
3 Department ofMechanical andIndustrial Engineering,
University ofZawia, Tripoli, Libya
4 Department ofComputer Science, Faculty ofInformation
Technology, Zarqa University, Zarqa13110, Jordan
5 Department ofMedicine, Faculty ofMedicine, Hashemite
University, Zarqa13133, Jordan
6 Artificial Intelligence andSensing Technologies (AIST)
Research Center, University ofTabuk, Tabuk71491,
SaudiArabia
7 Hourani Center forApplied Scientific Research, Al-Ahliyya
Amman University, Amman19328, Jordan
8 MEU Research Unit, Middle East University, Amman,
Jordan
9 Department ofElectrical andComputer Engineering,
Lebanese American University, Byblos13-5053, Lebanon
10 School ofComputer Sciences, Universiti Sains Malaysia,
11800GeorgeTown, PulauPinang, Malaysia
11 School ofEngineering andTechnology, Sunway University
Malaysia, 27500PetalingJaya, Malaysia
12 Computer Science Department, Al al-Bayt University,
Mafraq25113, Jordan