Advancing Cyberbullying Detection: A Hybrid Machine Learning and

Deep Learning Framework for Social Media Analysis

Bishal Shyam Purkayastha

, Md. Musﬁqur Rahman

, Md. Towhidul Islam Talukdar

and Maryam Shahpasand

Computer Science (Cyber Security), University of Staffordshire London, London, U.K.

Department of Computer Science and Engineering, Chittagong University of Engineering and Technology, Chittagong,

Bangladesh

Department of Mechanical Engineering, Chittagong University of Engineering and Technology, Chittagong, Bangladesh

Keywords:

Automated Cyberbullying Detection, Transformer-Based Models, Social Media Text Analysis.

Abstract:

Social media platforms have led to the prevalence of cyberbullying, seriously challenging the mental health of

individuals. This research is on how effectively different machine learning and deep learning techniques can

detect cyberbullying in online communications. Using two different tweet datasets obtained from Mandalay

and Kaggle, we developed a balanced framework for binary classiﬁcation. This research emphasizes compre-

hensive data preprocessing: text normalization and class balancing by random oversampling to increase the

dataset’s quality. Models used include several traditional machine learning classiﬁers: Random Forest, Extra

Trees, AdaBoost, MLP, and XGBoost, and advanced deep learning architectures such as Bidirectional LSTM,

BiGRU, and BERT. These results conﬁrm that deep learning models, especially BERT, yield outstanding per-

formance with an accuracy rate of 92%, hence showing the models’ capability in effectively detecting and

preventing cyberbullying through automated detection.

1 INTRODUCTION

Bullying is a deliberate act of aggression where indi-

viduals exploit their social or physical dominance to

harm others, often targeting those who are less pow-

erful. It manifests in various forms—verbal, physi-

cal, or social—and inﬂicts signiﬁcant emotional and

psychological suffering on victims. Cyberbullying,

an extension of this behavior into digital environ-

ments, has emerged as a pressing social concern with

the widespread use of the internet and social media

platforms. Studies reveal that approximately 37% of

young individuals in India have experienced cyber-

bullying, with 14% enduring chronic instances (Arif,

2021). This form of digital hostility manifests as ha-

rassment, intimidation, or public humiliation through

social media and other online channels, profoundly

affecting victims’ mental health, academic perfor-

mance.

The evolution of cyberbullying has closely followed

advancements in technology depicted in Figure 1.

In the 1990s, it began in internet forums and chat

rooms, where anonymity enabled verbal harassment

and rumor-spreading. The 2000s saw the rise of social

networks like Myspace and Friendster, amplifying the

impact of bullying through public humiliation and im-

personation. By the 2010s, anonymous messaging

apps such as Sarahah and Ask.fm further complicated

the issue, allowing bullies to attack without fear of ac-

countability. The late 2010s introduced deepfake ma-

nipulation, enabling bullies to create and share false,

humiliating content. More recently, the 2020s have

seen the emergence of AI-powered harassment, au-

tomating and intensifying bullying attacks, making

them harder to monitor and counteract.

As technology advances, so do the tactics employed

by cyberbullies, underscoring the urgent need for ef-

fective intervention mechanisms. Machine learning

(ML) and deep learning (DL) methods have emerged

as crucial tools in the ﬁght against cyberbullying, of-

fering the ability to analyze large volumes of text and

multimedia data to identify patterns of abusive behav-

ior (Bruwaene et al., 2020).

In response to these challenges, this study con-

ducts a comprehensive evaluation of ML- and DL-

based approaches for detecting and classifying cy-

348

Purkayastha, B. S., Rahman, M. M., Talukdar, M. T. I. and Shahpasand, M.

Advancing Cyberbullying Detection: A Hybrid Machine Learning and Deep Learning Framework for Social Media Analysis.

DOI: 10.5220/0013436200003929

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 2, pages 348-355

ISBN: 978-989-758-749-8; ISSN: 2184-4992

Figure 1: Evaluation of cyberbullying.

berbullying incidents. By leveraging two authentic

datasets, the research develops and validates a bi-

nary classiﬁcation framework that identiﬁes cyberbul-

lying occurrences within social media content. Rigor-

ous data preprocessing, feature engineering, and ad-

vanced transformer-based models are employed to as-

sess the strengths and limitations of various detection

methods, demonstrating their potential to mitigate the

growing issue of cyberbullying.

The main contributions of our work are:

1. Integrated two real-world datasets from Mandalay

and Kaggle, addressing class imbalances through

random oversampling to create a balanced binary

classiﬁcation framework.

2. Developed an effective data preprocessing and

feature engineering workﬂow, enhancing the qual-

ity and reliability of input data for model training

and evaluation.

3. Conducted an extensive evaluation of machine

learning and deep learning models, highlighting

BERT’s superior performance with an accuracy of

92%.

2 RELATED WORK

Cyberbullying detection has garnered signiﬁcant at-

tention in recent years, with diverse approaches lever-

aging machine learning (ML) and deep learning (DL)

methods to identify harmful behaviors in online com-

munications. In (Balakrishnan et al., 2020), proposed

an automatic detection method that utilizes the psy-

chological characteristics of Twitter users for feature

extraction and classiﬁcation. The authors tested Ran-

dom Forest (RF) and J48 classiﬁers using a dataset

of 5,453 tweets, achieving promising results. Simi-

larly, (Yadav et al., 2020) employed the BERT model,

a transformer-based architecture, which effectively

generated contextual embeddings and achieved reli-

able outcomes in detecting cyberbullying.

In paper (Dalvi et al., 2020), explored traditional ap-

proaches using TF-IDF vectorization combined with

Naive Bayes (NB) and Support Vector Machines

(SVM) for tweet classiﬁcation, where SVM outper-

formed NB with an accuracy of 71.25%. In addition

(Hani et al., 2019), implemented a supervised learn-

ing method incorporating n-gram language models

and sentiment analysis, emphasizing feature extrac-

tion techniques to improve performance. In (Al-Ajlan

and Ykhlef, 2018), introduced Optimized Twitter Cy-

berbullying Detection (OCDD), leveraging convolu-

tional neural networks (CNNs) and Glove embed-

dings, further enhanced by meta-heuristic optimiza-

tion methods.

Recent works emphasize integrated frameworks and

ensemble methods. For example, (Unnava and

Parasana, 2024) compared various ML techniques

such as NB, k-Nearest Neighbors (kNN), Decision

Tree (DT), RF, and SVM, demonstrating notable per-

formance improvements with feature engineering. In

(Atoum, 2020), highlighted that utilizing n-grams in

NB outperformed SVM for Twitter-based datasets.

Moreover (Mehendale et al., 2022), explored offen-

sive language detection in multilingual datasets using

Natural Language Processing (NLP) and ML tech-

niques. Similarly, (Yuvaraj et al., 2021) proposed

an integrated model incorporating user context, psy-

chometric properties, and CB classiﬁcation. Emerg-

ing methods focus on addressing the challenges posed

by cyberbullying detection, including data imbalance,

context analysis, and multimodal approaches. In (Roy

and Mali, 2022), proposed a deep transfer learn-

ing model for image-based cyberbullying detection,

achieving an accuracy of 89%, though with limited

attention to textual data.

Despite advancements, critical gaps remain, including

the development of scalable, real-time solutions and

methods that generalize across modalities and lan-

guages.

3 CLASSIFICATION OF

CYBERBULLYING

Cyberbullying encompasses various forms of harass-

ment, each employing distinct tactics and method-

Advancing Cyberbullying Detection: A Hybrid Machine Learning and Deep Learning Framework for Social Media Analysis

349

ologies depicts in Figure 2. Direct cyberbullying

includes abusive messages, threats, and dissemina-

tion of false information. Cyberstalking involves

the intimidation of victims through persistent surveil-

lance and excessive communication. Flaming refers

to heated and offensive exchanges in public fo-

rums. Banning pertains to the exclusion of individuals

from online groups, often coupled with impersonation

through the creation of fake proﬁles. Outing and fraud

involve the unauthorized disclosure of personal in-

formation, while proxy cyberbullying leverages third

parties to harass victims. Lastly, catﬁshing describes

the emotional manipulation and exploitation of indi-

viduals using fabricated identities.

Figure 2: Classiﬁcation of cyberbullying based on the na-

ture and method of harassment.

Cyberbullying can be further categorized by the

methods and media employed. Verbal cyberbullying

uses harmful language, such as insults or threats via

messages or comments. Physical cyberbullying in-

volves unauthorized access to online accounts. Social

cyberbullying occurs on platforms where false infor-

mation is spread, or individuals are excluded. Psycho-

logical cyberbullying employs manipulation to inﬂict

emotional distress. Sexual cyberbullying entails the

dissemination of explicit content without consent, and

homophobic, racist, or religious cyberbullying targets

individuals based on identity, race, or beliefs, often

leveraging hateful language or discrimination.

4 METHODOLOGY

This research employs a systematic approach, begin-

ning with the collection of two open-source datasets

from Mandalay and Kaggle, followed by preprocess-

ing to enhance data quality. The processed dataset is

then subjected to machine learning (ML) algorithms

for cyberbullying detection. The workﬂow is depicted

in Figure 3.

4.1 Dataset Collection

Two open-source datasets were utilized: a multiclass-

labeled dataset from Mandalay - Cyberbullying

Tweets, (Mehendale et al., 2022) and a binary-labeled

dataset from Kaggle - Cyberbullying Classiﬁcation

(Kaggle, 2024). To create a binary-class hybrid

dataset, relevant classes were relabeled such that all

instances of cyberbullying were labeled as 1 and non-

cyberbullying instances as 0. Due to class imbal-

ance, random oversampling was employed, resulting

in a balanced dataset comprising 90,276 records. This

dataset was subsequently used to train ML models for

cyberbullying detection.

4.2 Dataset Preprocessing

Preprocessing steps included the removal of hyper-

links, punctuation, extra spaces, and stop words to

improve data quality. All text was converted to lower-

case to address case inconsistencies, and non-English

words were translated to English. Random over-

sampling was applied to balance the classes, ensur-

ing equal representation of cyberbullying and non-

cyberbullying instances. These steps enhanced the

dataset’s suitability for machine learning analysis.

5 BULLYING DETECTION

MODEL

The proposed framework for cyberbullying detection

integrates Natural Language Processing (NLP) and

Machine Learning (ML). The methodology is cate-

gorized into two major components: neural network-

based approaches and classical machine learning al-

gorithms.

5.1 Neural Network Approaches

5.1.1 Bidirectional Long Short-Term Memory

(BiLSTM)

The BiLSTM model processes sequences in both for-

ward and backward directions, enabling a comprehen-

sive understanding of contextual relationships in text

data. For a given input sequence X = [x

, x

, . . . , x

−→

= σ(W

−→

t−1

+ b

) (1)

←−

= σ(W

←−

t+1

+ b

) (2)

Here, σ is the activation function, W

and W

are

weight matrices, and b

is the bias term. The ﬁnal

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

350

Figure 3: Classiﬁcation of cyberbullying based on nature and method of harassment.

hidden state is the concatenation of forward and back-

ward states:

= [

−→

;

←−

] (3)

This bidirectional approach ensures that the model

captures dependencies from both past and future con-

texts, making it particularly effective for sentiment

analysis and text classiﬁcation tasks (Gada et al.,

2021).

5.1.2 Gated Recurrent Unit (GRU)

The GRU simpliﬁes Long Short-Term Memory

(LSTM) models by using fewer gates while maintain-

ing performance. The update and reset gates are com-

puted as follows:

= σ(W

t−1

+ b

) (4)

= σ(W

t−1

+ b

) (5)

The hidden state is updated as:

= z

⊙h

t−1

+ (1 −z

) ⊙

(6)

where

= tanh(W

+ r

⊙(U

t−1

)). GRUs are

computationally efﬁcient and work well for sequen-

tial data tasks like language modeling and named en-

tity recognition (Fang et al., 2021).

5.1.3 Bidirectional Encoder Representations

from Transformers (BERT)

BERT utilizes self-attention mechanisms to generate

context-aware embeddings by processing text bidirec-

tionally. For an input sequence X:

Q = XW

, K = XW

, V = XW

(7)

Attention(Q, K, V ) = softmax



√



V (8)

Here, Q, K, and V represent the query, key, and value

matrices, and d

is the dimensionality of the keys.

BERT employs masked language modeling (MLM)

and next sentence prediction (NSP) for pre-training,

enabling it to understand sentence relationships and

word context efﬁciently (Paul and Saha, 2022). It is

particularly powerful for tasks like sentiment analysis

and query answering.

5.2 Classical Machine Learning

Algorithms

5.2.1 Naive Bayes Classiﬁer

The Naive Bayes classiﬁer is a probabilistic approach

based on Bayes’ theorem. Given a feature vector X =

, x

, . . . , x

], the probability of class C is computed

as:

P(C|X) =

P(C)

∏

i=1

P(x

|C)

P(X)

(9)

The decision rule is:

∗

= arg max

P(C)

∏

i=1

P(x

|C) (10)

This classiﬁer assumes independence among features,

making it simple yet effective for text classiﬁcation

tasks, especially with well-engineered features (Paul

and Saha, 2022).

5.2.2 Random Forest

The Random Forest algorithm operates as an ensem-

ble of decision trees. For a dataset D, the classiﬁca-

tion decision is obtained by aggregating predictions

from T trees:

F(x) = majority vote(T

(x), T

(x), . . . , T

(x)) (11)

Advancing Cyberbullying Detection: A Hybrid Machine Learning and Deep Learning Framework for Social Media Analysis

351

Each tree is trained on a bootstrap sample of the

dataset, and feature selection during tree construction

introduces diversity. The majority vote determines the

ﬁnal class label (Raj, 2021).

5.2.3 XGBoost

XGBoost (Extreme Gradient Boosting) optimizes

gradient boosting by minimizing a regularized loss

function. Given the prediction ˆy

, the objective is:

L =

∑

i=1

l(y

, ˆy

) +

∑

k=1

Ω( f

) (12)

where Ω( f

) = γT +

λ∥w∥

, T is the number of

leaves, and w are the leaf weights. This algorithm is

particularly effective for structured data (Raj, 2021).

5.2.4 AdaBoost

AdaBoost combines weak classiﬁers to form a strong

classiﬁer. For each iteration t:



1 −e



(13)

where e

is the weighted error rate. The updated

weights for misclassiﬁed samples emphasize their im-

portance in subsequent iterations (Raza et al., 2020).

5.3 Feature Engineering

Feature engineering is crucial for distinguishing be-

tween cyberbullying and neutral content. Measurable

attributes such as lexical and syntactic features, de-

mographic details, and sentiment scores are incorpo-

rated. The sentiment score for a document is com-

puted as:

Sentiment Score =

∑

i=1

·s

(14)

where w

is the weight assigned to term i, and s

is the

sentiment score of term i. Effective feature selection

and generation enhance classiﬁcation performance by

ensuring the model focuses on relevant aspects of the

data.

5.4 Explainable AI (XAI)

Explainable AI (XAI) techniques, such as SHAP

(Shapley Additive Explanation) and LIME (Locally

Interpretable Model-Agnostic Explanation), are em-

ployed to interpret model decisions. The SHAP value

for a feature i is computed as:

∑

S⊆N\{i}

|S|!(|N|−|S|−1)!

|N|!

[v(S ∪{i}) −v(S)]

(15)

where S is a subset of features, N is the set of all fea-

tures, and v(S) is the model prediction for feature sub-

set S.

6 RESULTS AND DISCUSSION

This study provides a detailed comparative analysis

of machine learning (ML) and deep learning (DL)

classiﬁers for cyberbullying (CB) detection. The ex-

periments utilized hybrid datasets sourced from Man-

dalay and Kaggle to address the challenges associ-

ated with detecting diverse categories of cyberbully-

ing. Data preprocessing, feature engineering, and ad-

vanced classiﬁcation algorithms were applied to en-

sure comprehensive analysis.

6.1 Data Preparation and Evaluation

Metrics

The datasets underwent rigorous preprocessing steps

to enhance model performance. These steps included:

• Removal of punctuation, stop words, numbers,

and emojis.

• Spelling correction and language translation.

• Text normalization through compression manage-

ment and lowercase conversion.

Post preprocessing, the data was split into 70% train-

ing and 30% testing sets. Feature engineering meth-

ods such as term frequency-inverse document fre-

quency (TFIDF) and sentence embeddings were em-

ployed.

Evaluation metrics used for the models include:

Accuracy =

T P + T N

T P + FP + T N + FN

(16)

F1 Score = 2 ·

Precision ·Recall

Precision + Recall

(17)

Precision =

T P

T P + FP

(18)

Recall =

T P

T P + FN

(19)

6.2 Performance of ML and DL Models

Table 1 illustrates the performance metrics for six ML

classiﬁers. Among these, the Extra Trees classiﬁer

emerged as the best performer with a 90.24% accu-

racy and 90.21% F1 score. This model was optimized

using hyperparameters such as n estimators: 109,

learning rate: 0.1, max depth: 25, min samples split:

9, and min samples leaf : 1.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

352

Table 1: Performance Metrics of ML Classiﬁers.

Model F1 Score Accuracy Recall Precision

Random Forest 0.8831 0.8836 0.8836 0.8955

Extra Trees 0.9021 0.9024 0.9024 0.9112

AdaBoost 0.8024 0.8039 0.8039 0.8188

MLP 0.8873 0.8876 0.8876 0.8959

XGBoost 0.8397 0.8416 0.8416 0.8654

Gradient Boost 0.8036 0.8064 0.8064 0.8332

Table 2 highlights the performance of three DL

models. BERT achieved the highest accuracy of

91.36% and an F1 score of 91.24%, outperforming

other DL models. The results underscore the effec-

tiveness of BERT in capturing contextual relation-

ships in textual data.

Table 2: Performance Metrics of DL Models.

Model F1 Score Accuracy Recall Precision

BiLSTM 0.9084 0.9105 0.8605 0.9619

BiGRU 0.9053 0.9053 0.8779 0.9345

BERT 0.9124 0.9136 0.8723 0.9564

The confusion matrix for the Extra Trees classi-

ﬁer in Figure 4 demonstrates its ability to accurately

classify cyberbullying and non-cyberbullying events.

True positives (TP) and true negatives (TN) signiﬁ-

cantly outweighed false positives (FP) and false neg-

atives (FN), afﬁrming the model’s robustness.

Figure 4: Confusion Matrix for Extra Trees Classiﬁer.

Similarly, the ROC curve in Figure 5 highlights

the model’s efﬁciency, achieving an AUC score of

0.96, indicative of its high sensitivity and speciﬁcity

in distinguishing cyberbullying instances.

6.2.1 Interpretability and XAI Analysis

Model interpretability was enhanced using explain-

able AI (XAI) techniques like SHAP and LIME. Fig-

ure 6 illustrates the LIME-based interpretation for the

Figure 5: ROC Curve for Extra Trees Classiﬁer.

Extra Trees classiﬁer, highlighting feature contribu-

tions in predicting cyberbullying.

Figure 6: Local Interpretability with LIME: Contribution of

Words to the CB Class.

LIME identiﬁed words such as Muslim, terror-

ist, and Qur’an as key indicators for the CB class.

These insights enhance transparency and foster trust

in automated cyberbullying detection systems. Fig-

ure 7 further exempliﬁes the interpretability of the

classiﬁer by highlighting speciﬁc words, such as ig-

norant, reason, and ghetto, that inﬂuenced predic-

tions. Each panel demonstrates the contribution of

these words toward classifying instances as bullying

or not bullying. The visualizations emphasize the im-

portance of these features in understanding the classi-

ﬁer’s decision-making process.

6.3 Discussion

The advent of social networking platforms has revolu-

tionized communication, offering unparalleled oppor-

tunities for global interaction. However, these plat-

forms have also become a medium for cyberbully-

ing, where individuals are harassed, intimidated, or

Advancing Cyberbullying Detection: A Hybrid Machine Learning and Deep Learning Framework for Social Media Analysis

353

Figure 7: LIME-based Interpretability Analysis: Contribution of Words to Bullying and Not Bullying Classes.

harmed through electronic communication. This per-

vasive issue poses signiﬁcant risks to mental health

and social well-being, necessitating robust detection

and prevention mechanisms. This study investigates

the efﬁcacy of machine learning (ML) and deep learn-

ing (DL) models in addressing this critical challenge,

emphasizing their performance, limitations, and fu-

ture implications.

6.3.1 Key Findings and Performance Analysis

The results underscore the effectiveness of ML and

DL approaches in detecting cyberbullying on so-

cial media platforms. By employing algorithms

such as Random Forest, Extra Trees, MLP, XG-

Boost, AdaBoost, Gradient Boost, BiLSTM, BiGRU,

and BERT, a comprehensive evaluation was con-

ducted using accuracy, F1 score, precision, recall, and

area under the receiver operating characteristic curve

(AUC-ROC). Among the ML classiﬁers, Extra Trees

emerged as the most effective, achieving an accuracy

of 90.24% and an F1 score of 90.21%. For DL mod-

els, BERT outperformed others with 91.36% accuracy

and an F1 score of 91.24%, due to its bidirectional

contextual information processing.

6.3.2 Advantages of Deep Learning Models

Deep learning models, particularly BiLSTM, BiGRU,

and BERT, showed superior performance compared

to traditional ML algorithms. Their capacity to

capture sequential dependencies and contextual nu-

ances in textual data makes them particularly suit-

able for analyzing interactions on social media plat-

forms. BERT’s ability to leverage attention mecha-

nisms helped it detect subtle linguistic features, en-

hancing its differentiation between cyberbullying and

non-cyberbullying content. These attributes under-

score the potential of DL models in handling the com-

plexities of cyberbullying detection.

6.3.3 Challenges and Limitations

While the ﬁndings highlight signiﬁcant progress,

challenges remain. The dynamic nature of on-

line communication presents difﬁculties in accurately

identifying cyberbullying. Language nuances, cul-

tural differences, and adversarial behaviors, such as

coded language, often elude static models. Addition-

ally, reliance on labeled datasets introduces potential

biases, limiting the generalizability of models across

diverse contexts.

The ﬁndings of this study, underscores the impor-

tance of robust evaluation mechanisms, such as con-

fusion matrices and AUC-ROC analysis, in optimiz-

ing detection accuracy. Deep learning models, partic-

ularly BERT, showed exceptional performance, lever-

aging contextual understanding to address cyberbul-

lying complexities. However, challenges related to

language dynamics, adversarial behaviors, and ethi-

cal considerations remain. Addressing these issues

through innovative approaches and interdisciplinary

collaboration is key to creating a safer online envi-

ronment. Future efforts should prioritize multilingual,

multimodal detection systems and the integration of

explainability techniques to ensure transparency and

fairness, playing a crucial role in combating cyber-

bullying while upholding ethical standards.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

354

7 CONCLUSION

Cyberbullying has become a critical concern in the

digital age, impacting individuals and society on mul-

tiple levels. This study highlights the importance of

employing robust detection mechanisms to address

this growing issue effectively. Among the machine

learning classiﬁers, the Extra Trees algorithm outper-

formed traditional methods, achieving notable results

with an accuracy of 90.24%, precision of 91.12%, re-

call of 90.24%, and an F1-score of 90.21%. While

these results are signiﬁcant, deep learning models

demonstrated even greater efﬁcacy. In particular,

attention-based architectures and bidirectional neural

networks emerged as the most effective approaches,

with the BERT-based model achieving the highest

metrics: 91.36% accuracy, 93.45% precision, 87.23%

recall, and 91.24% F1-score. This underscores the

advantage of using neural networks to capture the nu-

anced and context-dependent nature of cyberbullying-

related text. Notably, our shallow neural network

framework offers a resource-efﬁcient alternative, re-

ducing the need for complex deep neural networks

while maintaining competitive performance.

Future research should explore hybrid and ensemble

methods to further improve detection accuracy and re-

silience. By focusing on these areas, researchers and

practitioners can develop more comprehensive and

scalable solutions to combat cyberbullying, ensuring

safer online environments for all users.

REFERENCES

Al-Ajlan, M. A. and Ykhlef, M. (2018). Optimized twitter

cyberbullying detection based on deep learning. In

2018 21st Saudi Computer Society National Computer

Conference (NCC), pages 1–5. IEEE.

Arif, M. (2021). A systematic review of machine learning

algorithms in cyberbullying detection: Future direc-

tions and challenges. Journal of Information Security

and Cybercrimes Research, 4(1):01–26.

Atoum, J. O. (2020). Cyberbullying detection through sen-

timent analysis. In 2020 International Conference

on Computational Science and Computational Intel-

ligence (CSCI), pages 292–297. IEEE.

Balakrishnan, V., Khan, S., and Arabnia, H. R. (2020). Im-

proving cyberbullying detection using twitter users’

psychological features and machine learning. Com-

puters & Security, 90:101710.

Bruwaene, L. V., Houtte, E. V., and Backer, M. D. (2020).

Utilizing machine learning and deep learning methods

for detecting and analyzing cyberbullying. Computers

in Human Behavior, 108:106–116.

Dalvi, R. R., Chavan, S. B., and Halbe, A. (2020). De-

tecting a twitter cyberbullying using machine learn-

ing. In 2020 4th International Conference on Intelli-

gent Computing and Control Systems (ICICCS), pages

297–301. IEEE.

Fang, Y. et al. (2021). Gated recurrent units in natural lan-

guage processing. AI Research Letters.

Gada, X. et al. (2021). Bidirectional lstm for text process-

ing. NLP Journal.

Hani, J., Nashaat, M., Ahmed, M., Emad, Z., Amer, E., and

Mohammed, A. (2019). Social media cyberbullying

detection using machine learning. International Jour-

nal of Advanced Computer Science and Applications,

10(5):703–707.

Kaggle (2024). Cyberbullying tweets. Retrieved Septem-

ber 30, 2024, from https://www.kaggle.com/datasets/

soorajtomar/cyberbullying-tweets.

Mehendale, N., Shah, K., Phadtare, C., and Rajpara, K.

(2022). Cyberbullying detection for hindi-english

language using machine learning. Available at

SSRN 4116143, https://papers.ssrn.com/sol3/papers.

cfm?abstract

id=4116143.

Paul, D. and Saha, A. (2022). Bidirectional encoder repre-

sentations from transformers (bert). Machine Learn-

ing Applications.

Raj, K. (2021). Xgboost: An improved gradient boosting

algorithm. Data Science Advances.

Raza, S. et al. (2020). Adaboost for weak classiﬁers. En-

semble Learning Journal.

Roy, P. and Mali, P. (2022). A deep transfer learning model

for image-based cyberbullying detection in social net-

works. International Journal of Advanced Computer

Science and Applications, 13(5):45–51.

Unnava, S. and Parasana, S. R. (2024). A study of cyberbul-

lying detection and classiﬁcation techniques: A ma-

chine learning approach. Engineering, Technology &

Applied Science Research, 14(4):15607–15613.

Yadav, J., Kumar, D., and Chauhan, D. (2020). Cyberbul-

lying detection using pre-trained bert model. In 2020

International Conference on Electronics and Sustain-

able Communication Systems (ICESC), pages 1096–

1100. IEEE.

Yuvaraj, N. et al. (2021). Nature-inspired-based approach

for automated cyberbullying classiﬁcation on multi-

media social networking. Mathematical Problems in

Engineering, 2021:1–12.

Advancing Cyberbullying Detection: A Hybrid Machine Learning and Deep Learning Framework for Social Media Analysis

355