Received 15 April 2024, accepted 22 May 2024, date of publication 28 May 2024, date of current version 6 June 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3406595

Deep Learning Algorithms for Cyber-Bulling

Detection in Social Media Platforms
MOHAMMED HUSSEIN OBAIDA 1,2 , SALEH MESBAH ELKAFFAS3 ,
AND SHAWKAT KAMAL GUIRGUIS2
1 College of Science, Al-Nahrain University, Jadriya, Baghdad 64074, Iraq
2 Department of Information Technology, Institute of Graduate Studies and Research, Alexandria University, Alexandria 21544, Egypt
3 College of Computing and Information Technology, Arab Academy for Science, Technology and Maritime Transport, Alexandria 19838, Egypt
Corresponding author: Mohammed Hussein Obaida ([email protected])

ABSTRACT Social media platforms are among the most widely used means of communication. However,
some individuals exploit these platforms for nefarious purposes, with ‘‘cyberbullying’’ being particularly
prevalent. Cyberbullying, which involves using electronic means to harass or harm others, is especially
common among young people. Consequently, this study aims to propose a model for detecting cyberbullying
using a deep learning algorithm. Three datasets from Twitter, Instagram, and Facebook were utilized to
predict instances of bullying using the Long Short-Term Memory (LSTM) method. The results obtained
revealed the development of an effective model for detecting cyberbullying, addressing challenges faced
by previous cyberbullying detection techniques. The model achieved accuracies of approximately 96.64%,
94.49%, and 91.26% for the Twitter, Instagram, and Facebook datasets, respectively.

INDEX TERMS Cyberbullying, deep learning, LSTM, social networks.

I. INTRODUCTION social networks, often leads individuals to exhibit more

With the advent of online social networks, the widespread aggressive behavior [3]. Additionally, the growing trend of
accessibility of information and communication technol- expressing opinions freely online has also contributed to
ogy, as well as the prevalent utilization of computers and the spread of hate speech. Given the detrimental impact of
smartphones, individuals on the Internet are experiencing such prejudiced communication on society, governments and
heightened degrees of freedom of speech. Furthermore, users social media platforms stand to benefit from the utilization of
of social media platforms frequently have the ability to tools for the detection and prevention of hate speech [4].
conceal their identities, thereby enabling the exploitation Deep Learning (DL) is a method employed in the field
of various functionalities. The issue of offensive language of machine learning that enables the performance of unsu-
has become prominent in the domain of social networking. pervised learning using un labeled data. In the areas of
Offensive language refers to any form of communication that data mining and text classification, a variety of research
displays abusive conduct with the purpose of causing harm to investigations have applied DL methodologies to predict and
others. Various types of abusive language can be observed classify events such as detecting hate speech and categorizing
on social networking platforms, including sexism, racism, opinions. Various types of Deep Learning Networks include
cyberbullying, hate speech, and toxic remarks [1]. Feed Forward Neural Networks, Deep Belief Networks, Con-
Hate speech has become increasingly prevalent in both volutional Neural Networks (CNN), Restricted Boltzmann
face-to-face interactions and online communication in recent Machines, Recurrent Neural Networks (RNN), and Stacked
years [2]. Various factors play a role in this phenomenon. Denoising Autoencoders [5].
Firstly, the anonymity offered by online platforms, especially Numerous approaches have been explored in the quest
to detect hate speech, including conventional classifiers,
The associate editor coordinating the review of this manuscript and neural network-driven classifiers, or a combination of both
approving it for publication was Nuno M. Garcia . techniques. Multi-layer Perceptron (MLP), support vector
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 12, 2024 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 76901
 M. H. Obaida et al.: Deep Learning Algorithms for Cyber-Bulling Detection

machines (SVM) and extreme gradient boosting (XGB) the enhancement in performance compared to standard
are commonly used as classifiers in this domain, where references.
the application of vectorized representations of textual
data is typically essential. Commonly employed approaches
include the utilization of bag-of-word models in conjunction II. RELATED WORKS
with TF-IDF (term frequency-inverse document frequency). Several methods offer models for cyberbullying detection.
The progress in embedding methodologies rooted in deep In the study of [8], the model is suggested to provide a dual
learning has brought forth tools like Fast Text, Glove, definition of cyberbullying by utilizing a creative CNN idea
word2vec, and transformer-based approaches, which have for content analysis as well as a dishonest method to deal
been employed to acquire more complex representations. with providing the arrangement with less accuracy. When
These embedding-based representations, pre-trained using compared to other studies, the collected data are proven to
representation learning tools, can be utilized with both tra- provide superior precision and categorization.
ditional and sophisticated classifiers. This expands the array A systematic review of n=186 entries from internet data-
of techniques available for identifying hate speech, providing banks was published by [9]. In this article, 10 reviews of
a diverse set of potential solutions applicable to various real- the literature have been chosen to assess and debate the data
world situations [6]. regarding the effectiveness of ML in preventing cyberbully-
The increase in cyberbullying on social networking sites ing. To predict cyberbullying, most models take advantage of
and the diversity of its forms has resulted in negative effects content-based features. While the most prevalent algorithms
on the victim. The negative effects that appeared on the are support vector machines, naive Bayes, and convolutional
victims after being exposed to cyberbullying are many, such neural networks, the majority of these traits are based on text
as negative effects on physical health and mental health like from social media posts. ML is a cutting-edge preventative
anxiety, depression, thinking, and low self-esteem, and some- technique that might enhance and combine adolescent edu-
times it led to suicide [27]. With the emergence of negative cation programs and serve as the foundation for the creation
effects and the increase of bullying on social media sites, of technology-based automated screening methods.
it has become necessary to find a solution to reduce and According to studies by [10], a technique to detect cyber-
prevent the phenomenon of cyberbullying [28]. bullying was created using fuzzy logic, in which the two
In a comparative study published by [29] for cyberbullying users’ communication is continuously observed, and each
detection in social media for the last five years, presented a message’s emotional content is identified. Depending on their
group of previous studies that used machine learning and deep emotions, each user’s behaviour is classified as either decent
learning algorithms in good attempts to detect and classify or bullying. The user’s account is automatically terminated
the phenomenon of cyberbullying. They concluded from their and reported if the amount of observed bullying exceeds a
observations of the results of related works for achieving predetermined threshold value. They concluded that if used
better results in future research.is recommended to use deep alongside social networking sites, it could be a helpful tool
learning algorithms (the BiLSTM classifier and BERT) while for avoiding online harassment. The created algorithm can
in the case of using machine learning algorithms, prefer use also be used for surveillance and studying human behaviour.
SVM and NB as classifiers. A novel pre-trained BERT model was developed by [11]
From the survey of the previous studies a number of limi- and assessed using two social media datasets. One dataset
tation has been addressed such as multi-class cyberbullying featured a comparatively small network layer at the top func-
categories, experiments with larger datasets for aggression tioning as a classifier, while the other dataset had a larger
detection, datasets from multiple social media platforms not network layer at the top serving as a classifier. The primary
taken in consideration. So the primary aim of this research objective of the study was to detect instances of cyberbullying
was to develop a detection model that enhances the per- on various social media platforms. When compared to earlier
formance of classifiers on a large generic dataset through methods, this one works better in terms of dimensions and
combining feature extraction techniques. The study presents training the model.
the creation of an LSTM deep learning detection model capa- A study [12] presented a new model known as DEA-
ble of identifying cyberbullying content in user comments RNN, which combines Elman-type recurrent neural networks
across three distinct social media platforms in real-time. The (RNNs) with a refined dolphin allocation algorithm (DEA).
experimental setup for the new datasets closely follows the A dataset of 10,000 tweets was used for evaluation, com-
methodology outlined in a prior publication [7]. in an attempt paring the model’s performance with various advanced
to solve the limitation, allowing us to examine how well the algorithms such as bidirectional long short-term memory (Bi-
suggested model performs on the chosen datasets and how LSTM), RNN, SVM, multinomial naive Bayes (MNB), and
adaptable it is to other datasets. Using other datasets makes random forests (RF). Results from the experiments indicated
the model more flexible in its detection of bullying. numer- that DEA-RNN outperformed all other methods across dif-
ous experiments were conducted with a wide range of time ferent scenarios, achieving an average accuracy of 90.45%,
steps utilizing a dataset from the real world, while adhering precision of 89.52%, recall of 88.98%, F1-score of 89.25%,
to a time-conscious evaluation technique that demonstrates and specificity of 90.94%.

76902 VOLUME 12, 2024

M. H. Obaida et al.: Deep Learning Algorithms for Cyber-Bulling Detection

A study conducted by [13] aimed to develop a model

capable of identifying cyberbullying in Bangla and Roman-
ized Bangla texts through the application of various machine
learning and deep learning algorithms. The study utilized
three distinct datasets sourced from social media, compris-
ing 5000 Bangla texts, 7000 Romanized Bangla texts, and
12000 texts in a mixed language format. These datasets
underwent preprocessing before being employed to train
machine learning and deep learning algorithms. The research
assessed the model performances through a comparative anal-
ysis involving metrics such as accuracy, precision, recall,
f1-score, and ROC area. The results revealed that the deep
learning algorithm CNN exhibited the highest performance,
achieving an accuracy of 84% on the Bangla dataset. In con-
trast, the machine learning algorithm multinomial naive
Bayes surpassed others by achieving 84% accuracy on the
Romanized Bangla dataset and 80% accuracy on the com-
bined dataset.
Researchers from [14] have illustrated that utilizing the
TF-IDF vectorizer with Farasa NLTK, SVM exhibited supe-
rior performance in cyberbullying classification when con-
trasted with the NB classifier. This was observed even when FIGURE 1. Architecture of the proposed models.

employing varying parameters for the ngram range and addi-

tional feature extraction like BoW. The findings indicated that
SVM continued to surpass NB in the identification of cyber- network API that integrates with the Tensor Flow open-source
bullying content, achieving an accuracy rate of 95.742%. machine learning framework [18]. This research utilized the
Study [15] implemented a sentiment detection system Natural Language Toolkit (NLTK), which is widely popu-
which incorporated Recurrent Neural Networks (RNN) for lar and recognized as one of the prominent NLP libraries
text evaluation and Convolutional Neural Networks (CNN) within the Python ecosystem. PyCharm was used for the
for image analysis. The textual information was collected entire implementation process. Through numerous experi-
using the Twitter API.. The effectiveness and functionality ments, the optimum conditions and results were established.
of the models were demonstrated, achieving an accuracy The proposed algorithm consists of four different steps of
ranging from 0.951 to 0.911. Additionally, F-measure scores cyberbullying detection that are applied in the state. The
of 0.910 and 0.890 were recorded for RNN and CNN, respec- subsections that follow describe and illustrate the main steps
tively. depicted in Fig. 1.
After examining the outcomes documented in the study
of [16], it was observed that linear SVC and SGD (stochas- B. DATA SETS
tic gradient classifier) demonstrated superior performance To identify cyberbullying textual content, the Hate-speech
in the classification and prediction of bullying messages datasets were used in the experiments of this study, including
in Hinglish. Furthermore, these algorithms required less the Twitter, Instagram, and Facebook datasets. Twitter is
time for training and predicting compared to other methods a widely-used platform for microblogging and social net-
like support vector machines, logistic regression, K nearest working, enabling users to share and engage with short
neighbour, random forest, bagging classifiers, decision tree messages known as ‘‘tweets.’’ The dataset used is available
classifiers, multinomial classifiers, and the AdaBoost classi- online at https://data.world/thomasrdavidson/hate-speech-
fier. and-offensive-language; this dataset includes 24783 tweets
A combination model of a convolutional neural network labelled as neither (1430), offensive language (19190), or hate
(CNN) and long short-term memory (LSTM) using the Kears (4163). The second dataset is Instagram, which is known as
Functional API was proposed by [17] After training the a social network built on media where users may upload,
model, the image-based prediction has an accuracy of 86%, like, and comment on pictures. One of the top five networks
and the text-based prediction has an accuracy of 85%. with the highest proportion of users reporting experiencing
cyberbullying is said to be this one [19]. It is available
at https://github.com/nurindahpratiwi/dataset-hate-speech-
III. METHODOLOGY instagram/blob/master/572-hate-speech-dataset.csv and con-
A. PROPOSED MODEL tains 572 Instagram posts labelled as neither (231), offensive
The model was implemented and trained as a sequential language (55) or hate (286). The third dataset specified for
type using Keras and Python. Keras is an advanced neural Facebook is available online at https://github.com/stjordanis/

VOLUME 12, 2024 76903

 M. H. Obaida et al.: Deep Learning Algorithms for Cyber-Bulling Detection

Dynamically-Generated-Hate-Speech-Dataset/blob/main/ categorization and forecasting. Such networks determine the

2020-12-31-Dynamically-Generated-Hate-Dataset-entries- relevance of information to retain or discard, as well as
v0.1.csv, which includes 40623 Facebook comments labelled the allocation of information to different neurons. These
as neither (5441), offensive language (16683), or hate networks employ backpropagation and a gated mechanism
(18499). in their operations. In the realm of a fundamental LSTM
Each dataset was in a CSV file form comprising two network, crucial elements include the input (it ), output (ot ),
columns, in which each row consists of a comment or a and forget gate (ft ), which are defined through mathematical
group of comments from a conversation and a corresponding equations. This study utilizes four concealed layers consisting
label, which for the used hate dataset was [0, 1, 2], with of 128 units, 64 units, 32 units, and 3 units respectively to effi-
1 denoting hate speech, 2 denoting offensive language, and ciently classify the remarks. The mathematical expressions
0 denoting neither, to train the categorized consideration are presented as follows:
system. By using the aforementioned notation, it is now
it = σ (Wi . ht−1, xt + bi )
 
(1)
possible to characterize the hate identification problem as the
ft = σ (Wi . ht−1, xt + bf )
 
process of learning how to use context, structural, temporal, (2)
ot = σ Wi . ht−1, xt + bo
 

and social information to determine whether a string of words (3)
or comments is relevant to cyberbullying.
the symbol h is utilized to characterize the condition of the
Data are noisy; thus, preprocessing is necessary to improve
input, with ht-1 representing the present state and ht-1 the
the quality of the extracted data, which includes eliminating
past state, while xt signifies an input text. The parameters for
all emoticons, folksonomies, slang terms, and stop words that
each gate are indicated as W and b, respectively. Within this
are not necessary for the study’s objectives. Tokenization,
framework, s denotes the activation function, specifically the
stemming lemmatization, and data cleansing were employed.
rectified linear unit (ReLU) in the proposed model. The ReLU
The text is tokenized by the Keras tokenizer, which then
function, unlike sigmoid and tanh, leads to sparse activation
converts it into a vector with binary coefficients for each
of neurons, implying that a neuron may not activate every
token based on the word count. The same length for all
time and could have a value of zero at certain instances.
sequences, ‘‘Pad sequences’’ was used. The dataset with
The activation function under discussion is recognized for
the clean sentence is then produced to run and forecast the
its efficacy in classification tasks. Referred to as the Recti-
indicated model.
fied Linear Unit (ReLU), this particular non-linear activation
function has gained popularity within the domain of deep
C. WORD EMBEDDING learning. A key benefit of ReLU is its unique characteristic
After the data preprocessing is completed, the model con- of not activating every neuron simultaneously, distinguishing
struction and training stage begins. The word embedding it from other activation functions. As a result, neurons will
technique is used to generate LSTM models. Word embed- remain activated until the outcome of the linear transfor-
dings are a set of methods for producing numerical repre- mation falls below 0. The training process of these models
sentations of text. Word embedding has the striking property involves the employment of a categorical cross-entropy loss
of producing identical representations for words with sim- function and utilizes backpropagation in combination with
ilar semantic meanings. This ability enables a machine to the Adam optimizer. The Adam optimizer is particularly
interpret the text as having meaning rather than just treating suitable for scenarios with a significant volume of data or
it as a collection of random integers. With random initial parameters due to its enhanced computational efficiency,
weights, the embedding layer of Keras was used to train the reduced memory usage, enhanced memory efficiency and
embeddings as a component of the network. This particular resistance against distortion during diagonal resizing are key
embedding option (Keras) stands out since it is task-specific. features of LSTM models, which are designed to accurately
By including semantics instead of just features obtained from categorize encoded documents into cyberbullying or non-
raw text, word embedding enhanced the suggested model’s cyberbullying.
level of accuracy in comparison to other conventional detec-
tion approaches [20]. E. PERFORMANCE EVALUATION
The ultimate outcome will be produced depending on the
D. LONG SHORT-TERM MEMORY (LSTM) ALGORITHM comprehensive categorization and forecasting. The effective-
Long-term dependencies in the data can be learned using ness of this suggested method is assessed through various
recurrent neural network (RNN) types called LSTM (long metrics such as the Accuracy of the classifier, which indicates
short-term memory) networks. In terms of information reten- the classifier’s capability to accurately predict the class label.
tion, long short-term memory networks (LSTMs) outperform Similarly, the predictor’s accuracy pertains to how accurately
RNNs. The vanishing gradient descent issue that standard a specific predictor can anticipate the value of the predicted
RNNs experience is resolved by LSTMs. Due to their exten- attribute for new data.
sive memory capabilities, Long Short-Term Memory (LSTM) TP + TN
Accuracy = (4)
networks are widely sought after for applications like text TP + TN + FP + FN

76904 VOLUME 12, 2024

M. H. Obaida et al.: Deep Learning Algorithms for Cyber-Bulling Detection

Precision can be described as the ratio of true positives to the TABLE 1. Performance of the proposed model in terms of the various
data set split ratios.
sum of true positives and false positives.
TP
Precision = (5)
TP + FP
Recall, in the context of information retrieval, refers to the
ratio of correctly retrieved results to the total number of
results that should have been retrieved. Within the realm of
binary classification, recall is alternatively known as sensi-
tivity. It can be interpreted as the likelihood that a pertinent
document will be brought back by a search query.
TP
Recall = (6) Similar results were obtained by [17]. The researchers indi-
TP + FN
cated that the model has been trained using 4,590,756 param-
The F measure, also known as F1 score or F score, quantifies eters, with distinct input and output specifically designed
the accuracy of a test by calculating the weighted harmonic for textual data. The model attains 85% accuracy by using
mean of the test’s precision and recall. sequentially dense LSTM layers. Additionally, the workers
2∗ TP in the research study conducted by [24] utilized an 80:20
F − measure = (7) ratio for dividing their data and indicated that their model
2∗ TP + FP + FN
consistently performed well in the early detection of cyber-
F. RESULTS AND DISCUSSION bullying. They utilized various characteristics to differentiate
The laptop used for all experiments was a DELL 3000 series, between positive and negative cases, setting low thresholds
Core (TM) i5-5200u, equipped with a Processor: 2.20 GHz to facilitate early detection and utilizing simpler features like
Intel (R), 12.0 GB of RAM and a 64-bit operating system. profile owner traits for the negative model. A study of [21]
The models were run multiple times across various epochs introduced an innovative approach to detecting cyberbully-
to ensure consistent evaluation parameters. To validate and ing, incorporating three advanced deep learning structures: a
ensure the reliability of the results, the performance of the multichannel architecture with BiGRU, a transformer block,
cyberbullying detection scheme was assessed using metrics and CNN models. The effectiveness of their methodology
such as accuracy, precision, recall, and f1 measure on a was evaluated, with the results of the experiments highlight-
dataset with minimal skewness. ing the importance of their strategy in categorizing short
Data splitting refers to the process of dividing the available messages (tweets). When the dataset was divided into 75%
data into two segments, typically for the purpose of cross- for training and 25% for testing, an accuracy rate of approxi-
validation. One portion is dedicated to developing a predictive mately 88% was attained.
model, while the other is utilized to evaluate the model’s Once the model is defined, it undergoes a compilation
performance. The process of splitting data into training and process in order to enable the execution of the Keras backend,
testing sets is a pivotal stage in the assessment of data mining which is based on Tensorflow. This compilation of the model
models. Typically, a significant portion of the data is desig- includes the incorporation of optimizers, loss functions, and
nated for training purposes, with a smaller fraction set aside metrics. Optimizers play a crucial role in updating the weights
for testing. of the model during the training phase. An important param-
In this research, the data was partitioned into various ratios eter utilized in this process is the number of epochs, the
for training and testing purposes, including 80% training and frequency at which the model encounters the training dataset
20% testing, 70% training and 30% testing, 60% training and is denoted by the number of epochs. This parameter acts
40% testing, and 50% training and 50% testing. The sets as a hyperparameter that specifies the number of iterations
were constructed by utilizing a randomized array, enabling the learning algorithm will go through the complete training
the model to adapt across various data samples. Additionally, dataset. During each epoch, all samples in the training dataset
this methodology aids in revealing the model’s dependability are utilized to adjust the internal model parameters. It is
and the uniformity of outcomes through multiple iterations. important to note that an epoch consists of at least one batch.
The random state is produced through numpy.random to facil- The accuracy of the model exhibited improvement with
itate random selection during the division of data, ensuring each subsequent epoch. However, it reached a point of sta-
consistent and replicable splits. bility after 40 epochs, as demonstrated in Table 2.
The obtained accuracy for the split ratio of the dataset was The model loss during different epochs for the used
at its best at 80% training and 20% testing. 96.64, 95.49, datasets is shown in Fig (2), Fig (3), Fig (4). The
and 91.42 for Tweets, Instagram, and Facebook, respectively. cross-entropy loss that was evaluated throughout various
as illustrated in Table 1, which revealed that the performance epochs within the configuration demonstrated effective con-
of the suggested method decreased as the number of trained vergence, suggesting an optimal level of performance for the
testers decreased in the latest group of research studies. model.

VOLUME 12, 2024 76905

 M. H. Obaida et al.: Deep Learning Algorithms for Cyber-Bulling Detection

TABLE 2. Model implementation over several epochs.

FIGURE 3. Accuracy and loss of LSTM model for the dataset2 (Instagram).

study by researchers [21]. Multichannel deep learning tech-

niques were employed using three models: the transformer
block, convolutional neural network, and bidirectional gated
recurrent unit (BiGRU), to classify Twitter comments into
aggressive and non-aggressive groups. Evaluation of perfor-
mance was conducted based on accuracy and compared to
various other machine learning algorithms known for their
FIGURE 2. Accuracy and loss of LSTM model for the dataset1(Twitter).
effectiveness in natural language processing tasks such as
CNN, BiLSTM, transformer, linear SVC, bagging, LR, RF,
and ET. The researchers reported that their approach out-
In the results obtained by [20], it was shown that cyberbul- performed others when the data was divided into 75% for
lying detection by LSTM achieved an accuracy of 72%, and training and 25% for testing, achieving an accuracy rate of
this performance increased from 65% to 72% after 10 epochs, around 88%.
while machine learning classifiers only achieved an accuracy On the contrary, a study [22] was carried out to investi-
of less than 70%. Finally, according to [25] they evaluated the gate the efficiency and efficacy of deep learning algorithms
LSTM model and revealed that LSTM was a viable method in detecting instances of abuse in social discourse. The
for identifying comments that were aggressive online, with study utilized four different deep learning models, namely
an accuracy of 93.84% and an f score of 0.94. LSTM, Bidirectional Long Short-Term Memory (BLSTM),
The authors in the study [26] claimed that the LSTM model Gated Recurrent Units (GRU) and Recurrent Neural Net-
shows a decreasing pattern in relation to the precision metric; work (RNN), to conduct experiments. The results indicated
consequently, as the number of epochs increases, the accuracy accuracies of 82.18%, 81.46%, 80.86%, and 81.01% respec-
level tends to settle at a lower value compared to its initial tively. Another study [23] found that deep learning methods
stage. Their findings were akin to the outcomes achieved in can effectively address cyberbullying in languages such as
our research with regards to epoch data. Roman and Urdu. Furthermore, advanced pre-processing
The performance of the proposed system was assessed in techniques involving lexicons and resources were imple-
comparison to an earlier researcher’s work as detailed in a mented, with the most optimal performance achieved using

76906 VOLUME 12, 2024

M. H. Obaida et al.: Deep Learning Algorithms for Cyber-Bulling Detection

TABLE 4. Comparison between proposed model and other studies in

terms of execution time.

citations as discussed in our prior publication [7]. The dura-

tion taken for one complete cycle of the dataset in question
is detailed in Table 4 within this investigation, which showed
that the required times were 493 sec, 21 sec, and 814 sec for
Twitter, Instagram, and Facebook, respectively.
The obtained results for the execution time were close to
those mentioned in other studies, such as the study of [12].
In contrast to prior deep learning architectures such as
FIGURE 4. Accuracy and loss of LSTM model for the dataset 3 (Facebook). Bi-LSTM and RNN baseline models, the newly suggested
DEA-RNN model necessitates a reduced amount of training
TABLE 3. Comparison between proposed model and previous studies in time. Specifically, the training duration for DEA-RNN was
terms of accuracy results. 248.52 seconds, as opposed to 349.1 and 274.31 seconds
for the baseline models that employed deep learning with
Bi-LSTM and RNN, respectively.
In a separate investigation conducted by [22], the
RNN-BiLSTM architecture was utilized. A sequential frame-
work was established which integrated a comprehensive set
of 2000 attributes within the embedding layer. The activation
function selected for the hidden layer (H1) was sigmoid.
Additionally, a binary cross entropy loss function was imple-
mented in conjunction with the Adam optimization method.
The model was subjected to numerous epochs, each taking
an average duration of 13 to 15 milliseconds per epoch.
Another research study (referenced as [25]) introduced the
RNN+LSTM model for identifying cyberbullying, indicat-
RNN-LSTM and RNN-BiLSTM, which attained accuracies ing an execution time of 78 minutes for this particular model.
of 85.5% and 85% for the aggression class. The study also
mentioned that upcoming research will concentrate on devel- IV. CONCLUSION AND FUTURE WORK
oping ensemble models to identify patterns of harassment and In this study, an efficient model to detect cyberbullying
hate speech. was established, inspired by the challenges faced by other
The results depicted in Table 3 illustrate that the newly cyberbullying detection techniques. Examples include feature
suggested approach exhibits superior performance in terms extraction by using deep learning algorithms. Additionally,
of accuracy compared to prior methodologies. This improve- the three different datasets that were used during this study
ment could be attributed to the LSTM’s ability to proficiently make the model more flexible in detecting bullies in compar-
capture term dependencies via its gating mechanism, as well ison with others. because the activation function influences
as its utilization of layers and gates to address the issue of how well the model understands the training data, ReLU was
vanishing gradients commonly encountered in a basic RNN. used instead of sigmoid for hidden layer activation, which
The final set of tests is performed to compare the difficulty improved LSTM performance. It is recommended to deter-
of the proposed system. In other terms, the utilization of mine the bullying by using the image and video and also using
computer resources by a program is explored using the same a multilingual dataset in future work.

VOLUME 12, 2024 76907

 M. H. Obaida et al.: Deep Learning Algorithms for Cyber-Bulling Detection

REFERENCES [16] K. Shah, C. Phadtare, and K. Rajpara, ‘‘Cyber-bullying detection in

[1] P. Fortuna and S. Nunes, ‘‘A survey on automatic detection of hate speech hinglish languages using machine learning,’’ Int. J. Eng. Res. Technol.,
in text,’’ ACM Comput. Surv., vol. 51, no. 4, pp. 1–30, Jul. 2019, doi: vol. 11, no. 5, pp. 439–447, May 2022.
10.1145/3232676. [17] V. V and H. P. D. Adolf, ‘‘Multimodal cyberbullying detection using hybrid
[2] FBI. (2015). Hatecrimestatistics. [Online]. Available: https://ucr.fbi.gov/ deep learning algorithms,’’ Int. J. Appl. Eng. Res., vol. 16, no. 7, p. 568,
hate-crime/ Jul. 2021, doi: 10.37622/ijaer/16.7.2021.568-574.
[3] P. Burnap and M. L. Williams, ‘‘Cyber hate speech on Twitter: An appli- [18] M. E. Kula, ‘‘Cyberbullying: A literature review on cross-cultural research
cation of machine classification and statistical modeling for policy and in the last quarter,’’ in Handbook of Research on Digital Violence and
decision making,’’ Policy Internet, vol. 7, no. 2, pp. 223–242, Jun. 2015. Discrimination Studies, F. Özsungur, Ed. Hershey, PA, USA: IGI Global,
[4] M. Wendling, The Year That Angry Won the Internet. London, U.K.: BBC 2022, pp. 610–630, doi: 10.4018/978-1-7998-9187-1.ch027.
Trending, 2015. [19] L. Cheng, R. Guo, Y. Silva, D. Hall, and H. Liu, ‘‘Hierarchical atten-
[5] B. Haidar, M. Chamoun, and A. Serhrouchni, ‘‘Arabic cyberbullying detec- tion networks for cyberbullying detection on the Instagram social net-
tion: Using deep learning,’’ in Proc. 7th Int. Conf. Comput. Commun. Eng. work,’’ in Proc. SIAM Int. Conf. Data Mining, 2019, pp. 235–243, doi:
(ICCCE), Sep. 2018, pp. 284–289, doi: 10.1109/ICCCE.2018.8539303. 10.1137/1.9781611975673.27.
[6] J. S. Malik, H. Qiao, G. Pang, and A. van den Hengel, ‘‘Deep learning for [20] E. Bashir and M. Bouguessa, ‘‘Data mining for cyberbullying and harass-
hate speech detection: A comparative study,’’ 2022, arXiv:2202.09517. ment detection in Arabic texts,’’ Int. J. Inf. Technol. Comput. Sci., vol. 13,
[7] M. H. Obaid, S. K. Guirguis, and S. M. Elkaffas, ‘‘Cyberbullying no. 5, pp. 41–50, Oct. 2021, doi: 10.5815/ijitcs.2021.05.04.
detection and severity determination model,’’ IEEE Access, vol. 11, [21] M. Alotaibi, B. Alotaibi, and A. Razaque, ‘‘A multichannel deep learn-
pp. 97391–97399, 2023, doi: 10.1109/ACCESS.2023.3313113. ing framework for cyberbullying detection on social media,’’ Electronics,
[8] V. Balakrishnan, S. Khan, and H. R. Arabnia, ‘‘Improving cyberbully- vol. 10, no. 21, p. 2664, Oct. 2021, doi: 10.3390/electronics10212664.
ing detection using Twitter users’ psychological features and machine [22] C. Iwendi, G. Srivastava, S. Khan, and P. K. R. Maddikunta, ‘‘Cyberbully-
learning,’’ Comput. Secur., vol. 90, Mar. 2020, Art. no. 101710, doi: ing detection solutions based on deep learning architectures,’’ Multimedia
10.1016/j.cose.2019.101710. Syst., vol. 29, no. 3, pp. 1839–1852, Jun. 2023, doi: 10.1007/s00530-020-
[9] J. D. Angelis* and G. Perasso, ‘‘Cyberbullying detection through 00701-5.
machine learning: Can technology help to prevent Internet bullying?’’ [23] A. Dewani, M. A. Memon, and S. Bhatti, ‘‘Cyberbullying detection:
Int. J. Manage. Humanities, vol. 4, no. 11, pp. 57–69, Jul. 2020, doi: Advanced preprocessing techniques & deep learning architecture for
10.35940/ijmh.k1056.0741120. Roman Urdu data,’’ J. Big Data, vol. 8, no. 1, pp. 1–20, Dec. 2021, doi:
[10] S. Prashar and S. Bhakar, ‘‘Real time cyberbullying detection,’’ Int. 10.1186/s40537-021-00550-7.
J. Eng. Adv. Technol., vol. 9, no. 2, pp. 5197–5201, Dec. 2019, doi: [24] M. F. López-Vizcaíno, F. J. Nóvoa, V. Carneiro, and F. Cacheda,
10.35940/ijeat.b4253.129219. ‘‘Early detection of cyberbullying on social media networks,’’
[11] J. Yadav, D. Kumar, and D. Chauhan, ‘‘Cyberbullying detection using pre- Future Gener. Comput. Syst., vol. 118, pp. 219–229, May 2021, doi:
trained BERT model,’’ in Proc. Int. Conf. Electron. Sustain. Commun. 10.1016/j.future.2021.01.006.
Syst. (ICESC), Jul. 2020, pp. 1096–1100, doi: 10.1109/ICESC48915.2020. [25] S. S. Shylaja, A. Narayanan, A. Venugopal, and A. Prasad, ‘‘Recurrent
9155700. neural network architectures with trained document embeddings for Flag-
[12] B. A. H. Murshed, J. Abawajy, S. Mallappa, M. A. N. Saif, ging cyber-aggressive comments on social media,’’ in Proc. Int. Conf. Adv.
and H. D. E. Al-Ariki, ‘‘DEA-RNN: A hybrid deep learning approach Comput. Commun. (ADCOM), 2018.
for cyberbullying detection in Twitter social media platform,’’ IEEE [26] D. A. Andrade-Segarra and G. A. Le, ‘‘Deep learning-based natural
Access, vol. 10, pp. 25857–25871, 2022, doi: 10.1109/ACCESS.2022. language processing methods comparison for presumptive detection of
3153675. cyberbullying in social networks,’’ Int. J. Adv. Comput. Sci. Appl., vol. 12,
[13] Md. T. Ahmed, M. Rahman, S. Nur, A. Islam, and D. Das, ‘‘Deployment of no. 5, pp. 796–803, 2021.
machine learning and deep learning algorithms in detecting cyberbullying [27] T. H. H. Aldhyani, M. H. Al-Adhaileh, and S. N. Alsubari, ‘‘Cyberbullying
in Bangla and romanized Bangla text: A comparative study,’’ in Proc. identification system based deep learning algorithms,’’ Electronics, vol. 11,
Int. Conf. Adv. Electr., Comput., Commun. Sustain. Technol. (ICAECT), no. 20, p. 3273, Oct. 2022, doi: 10.3390/electronics11203273.
Feb. 2021, pp. 1–10, doi: 10.1109/ICAECT49130.2021.9392608. [28] M. W. Habib and Z. N. Sultani, ‘‘Twitter sentiment analysis using different
[14] A. M. Alduailaj and A. Belghith, ‘‘Detecting Arabic cyberbullying tweets machine learning and feature extraction techniques,’’ Al-Nahrain J. Sci.,
using machine learning,’’ Mach. Learn. Knowl. Extraction, vol. 5, no. 1, vol. 24, no. 3, pp. 50–54, Sep. 2021.
pp. 29–42, Jan. 2023, doi: 10.3390/make5010003. [29] N. Haydar and B. N. Dhannoon, ‘‘A comparative study of cyberbullying
[15] M. Agbaje and O. Afolabi, ‘‘Neural network-based cyber-bullying and detection in social media for the last five years,’’ Al-Nahrain J. Sci., vol. 26,
cyber-aggression detection using Twitter text,’’ Babcock Univ., Jul. 2022, no. 2, pp. 47–55, 2023.
doi: 10.21203/rs.3.rs-1878604/v1.

76908 VOLUME 12, 2024