SMOTE-NGO of Stacking Ensemble Cyberbullying Detection of Feature and Model Parameter Optimization

Bullying is one form of aggressive behavior in which someone who has a lot of social or physical power usually abuses, threatens, or otherwise harms a weaker individual, who is generally a weaker person [1]. Bullying can take many different forms, including social (public humiliation), physical (hitting, kicking, pushing), or verbal (name-calling, threatening, taunting). Although they facilitate contact and the exchange of content, social media sites like Facebook, WhatsApp, Twitter, Instagram, and YouTube are frequently abused. The rapid increase in users across all age groups has led to a lack of awareness regarding proper usage and ethical behavior, resulting in harmful consequences such as anxiety, depression, and cyberbullying [2]. Children are particularly vulnerable to these harmful effects, and the lack of effective prevention measures only exacerbates the problem [3]. Victimization through cyberbullying was reported in the range of 13.99% to 57.5%, whereas the percentage of children and adolescents who experienced cyberbullying was within 6.0% to 46.3% [4, 5]. These incidents have been exacerbated due to the growth of social media use, frequently even without the users being aware of its purpose or ethical aspects. The advancement of technology, allowing the creation of new forms of bullying, has resulted in an increasing need for robust ways to identify and fight against cyberbullying [6]. Common cyberbullying channels include email, instant messaging, live chat rooms, text messages, social networks, and various websites [7, 8]. Cyberbullying behaviors include flooding (spamming content) and masquerading (impersonating victims) to incitement, trolling, hateful comments, humiliation, lying, posting embarrassing content, online exclusion, impersonation, trickery, and cyberstalking. These behaviors are increasingly prevalent among adolescents. Research and advancements in cyberbullying detection are urgently needed. Authorities can take the appropriate action to punish online harassers or stop them from posting anything that could incite cyberbullying if they identify instances of cyberbullying. In recent times, numerous studies have concentrated on the issue of cyberbullying on social media platforms and have proposed various methods to detect it. Therefore, it is crucial to conduct additional research on the English language and employ various techniques to obtain more precise results when identifying cyberbullying. Addressing class imbalance is a significant challenge in cyberbullying detection, as minority classes are often underrepresented in datasets. Various methods have been introduced to tackle this problem, such as oversampling, under sampling, and cost-sensitive learning. In this research, the Synthetic Minority Oversampling Technique (SMOTE) technique was utilized to balance the training data by creating synthetic samples for the minority class. Instead of duplicating existing data, SMOTE generates new samples through interpolation between existing instances. This method enhances class representation, improves the model's ability to generalize, and minimizes the likelihood of overfitting [9]. In this study, a hybrid cyberbullying detection framework is proposed. After splitting each dataset into 70% training and 30% testing sets, SMOTE is applied only to the training data to address class imbalance. The text is preprocessed by removing noise such as emojis, URLs, and special characters, and then converted into numerical features using Term Frequency-Inverse Document Frequency (TF-IDF). NGO is employed for feature selection and hyperparameter optimization of Random Forest (RF), AdaBoost, Support Vector Machine (SVM), and Naive Bayes (NB). The optimized models are integrated using a stacking ensemble. The framework is evaluated on five benchmark English cyberbullying datasets using 5-fold stratified cross-validation and performance metrics, including accuracy, precision, recall, F1-score, and Receiver Operating Characteristic–Area Under the Curve) ROC-AUC(. Experimental results indicate that the proposed stacking ensemble consistently outperforms the individual classifiers.

The following are this paper’s primary contributions:

1. This paper utilizes the NGO algorithm to achieve optimal Feature Selection.

2. NGO-based tuning for all four models (RF, Ada Boost, SVM, and NB).

3. Finally, integrate all optimized models into a unified ensemble framework.

This paper is broadly structured like this: Section 2 presents an overview of the literature review on cyberbullying detection and ensemble learning. Section 3 presents the proposed method and centralizes on the experimental results, data, and analyses the findings. Section 4 summarizes his conclusions and outlines the potential areas for future work.

1.png

A six-step process is outlined for cyberbullying detection on social media platforms:

1. The datasets are preprocessed using tokenization, lowercasing, text normalization, stop-word removal, and stemming.

2. TF-IDF is applied to convert the cleaned text into numerical feature vectors.

3. Each dataset is split into training (70%) and testing (30%) sets. SMOTE is then applied only to the training data to handle class imbalance.

4. NGO is used to select the most informative TF-IDF features and optimize the hyperparameters of RF, AdaBoost, SVM, and NB.

5. The top-performing optimized classifiers are combined using a stacking ensemble with AdaBoost as the meta-learner to perform binary classification. 6. The framework is evaluated using 5-fold stratified cross-validation and metrics including Accuracy, Precision, Recall, F1-score, ROC-AUC, PR-AUC, and matthews correlation coefficient (MCC).

3.1 Dataset collection

This study utilizes five benchmark English cyberbullying datasets collected from Kaggle, Twitter, and Wikipedia Talk Pages. The datasets cover different forms of abusive language, including aggression, toxicity, racism, sexism, and insults. In total, the combined datasets contain 312,115 labeled text samples [20]. Figure 2 illustrates the distribution of the selected datasets, and Table 2 summarizes their sources and sample sizes.

2.png

Figure 2. Cyberbullying datasets distribution

Table 2. Summary of cyberbullying datasets with sources and sample sizes

3.2 Data preparation

Data preparation is a key step in data mining. The several ways and steps for preparing data are all connected and work toward making data better for analysis. The type and amount of data mining work will define the preprocessing method [21]. To standardize and improve text data, you need to do a few things:

1. Tokenization divides text into smaller pieces known as tokens. Tokens might be words, phrases, or symbols. Preparing text for analysis is an important part of NLP [22].
2. Normalization removes extraneous data from large databases, such as noise tags, URLs, and hyperlinks. Preprocessing takes rid of problematic URLs and other data before undertaking text analysis or sentiment analysis [23].
3. Stop word: Words like “a,” “about,” “above,” and “again” are commonly removed from NLP preprocessing because they do not contribute help with text classification. Text data is improved by eliminating damaging or unnecessary phrases to enable precise cyberbullying identification [20]. The NLTK package for Python makes it easy to work with more than one language and swiftly gets rid of stop words from the text that has been obtained [24].
4. Stemming allows words to return to their basic forms by removing suffixes. For instance, “eaten” turns into “eat,” while “horses” turns into “horse.” Stemming reduces feature space, enhances classifier performance, and facilitates tasks such as machine learning and text categorization by reducing variations such as “connect, connects, connected, connecting” to a single root [25, 26].

3.3 Extracting features of cyberbullying

The process of extracting features is a key step in NLP tasks, including cyberbullying detection. To identify cyberbullying elements, to employ feature extraction techniques such as TF-IDF are employed, as described in the following sections.

3.3.1 Frequency representation of documents (TF-IDF)

TF-IDF is a feature extraction technique that weights words based on their frequency within a document and across the corpus. Eq. (1) illustrates how common terms in a document are given higher scores, while frequent words across all texts are down-weighted using inverse document frequency (IDF), resulting in modified term frequencies for training classifiers [27].

$T F-I D F=\left(T F_i * I D F_i\right)$               (1)

The frequency of word i in the document is denoted by the variable $T F_i$, while the inverse document frequency is denoted by $I D F_i$.

3.4 Optimizing hyperparameters and features

This section looks at NGO as a wrapper approach for Feature Selection and hyperparameter tweaking. It evaluates solutions based on model performance, such as classification accuracy or error rate, after iteratively training the model to optimize feature and input selection [28]. Wrapper approaches, such as NGO, require repetitive model training, but provide excellent accuracy in Feature Selection.

3.4.1 Northern goshawk optimization

This section introduces the NGO algorithm and its mathematical representation.

A. Inspiring and behavioral traits of the northern goshawk

The northern goshawk is a medium-sized bird from the Accipitridae family hunts rodents, rabbits, foxes, raccoons, and birds. Males in North America and Eurasia are smaller (780 g, 46-61 cm, wingspan 89-105 cm) than females (1220 g, 58-69 cm, wingspan 108-127 cm). They swoop down on prey with brief tail pursuits and hunt intelligently. Mathematics models were used to create the NGO algorithm [29].

B. Algorithm initialization

The population-based technique known as NGO was developed in response to the hunting habits of northern goshawks. With variable values recorded as vectors in a population matrix, every NGO member stands in for a solution. The search space is initialized at random at the beginning of the process, and the population is iteratively updated using the NGO criteria specified in [29].

$\left[\begin{array}{c}W_1 \\ \vdots \\ W_i \\ \vdots \\ W_N\end{array}\right]=\left[\begin{array}{ccccc}W_{1,1} & \cdots & W_{1, j} & \ldots & W_{1, m} \\ \vdots & \ddots & \vdots & \ddots & \vdots \\ W_{i, 1} & \cdots & W_{i, j} & \cdots & W_{i, m} \\ \vdots & \ddots & \vdots & \ddots & \vdots \\ X_{N, 1} & \cdots & W_{N, j} & \cdots & W_{N, m}\end{array}\right]$               (2)

Let W_i denote the ith candidate solution, and let W_i,j denote the value of the jth variable specified by ith candidate solution. The population of northern goshawks is denoted by W, with N being the population size and m the number of problem variables. Each individual in the population represents a potential solution to the optimization problem, enabling the evaluation of the objective function for every member. The objective function values across the population can be vectorized using Eq. (3) [29].

$\mathrm{Y}(\mathrm{x})=\left[\begin{array}{c}Y_1=Y\left(W_1\right) \\ \vdots \\ Y_i=Y\left(W_i\right) \\ \vdots \\ Y_N=Y\left(W_N\right)\end{array}\right]$             (3)

The vector of objective function values is denoted by Y, where Y_i represents the objective function value of the ith solution. The optimal solution is determined based on the objective function value: Lower values are preferred in minimization problems, while higher values are favored in maximization problems. Since objective function values are updated in each iteration, the best solution must be continuously tracked and updated accordingly.

C. Northern goshawk optimization mathematical modeling

The way northern goshawks hunt served as the model for an algorithm created by an NGO to update population data. It replicates two essential elements of the bird’s behavior: finding and catching prey and alternating between chasing targets or dodging dangers, all of which are done in two separate phases.

Phase 1: Prey identification (exploration): The northern goshawk quickly attacks its randomly chosen prey during the first stage of hunting. Similar to this, the NGO-inspired algorithm’s exploration capability is improved by using randomness in the selection of individuals from the search space. To find the most promising area, a worldwide search is conducted at the end of this stage. This stage is similar to the goshawk’s activity in identifying and pursuing prey. This initial stage is represented mathematically by Eqs. (4) and (5) [29].

$P_i=W_{K, i}=1,2, \ldots \ldots N, K=1,2, \ldots, i-1, i, i+1, \ldots \ldots, N$.             (4)

$W_{i, j}^{\text {new, } p_1}= \begin{cases}W_{i, j}+r\left(p_{i, j}-I W_{i, j}\right), & Y_{P_i}<Y_i \\ W_{i, j}+r\left(W_{i, j}-P_{i, j}\right), & Y_{P_i} \geq Y_i\end{cases}$             (5)

$W_i=\left\{\begin{array}{l}W_i^{\text {new }, p_1}, Y_i^{\text {new }, P_1}<Y_i \\ W_i, Y_i^{\text {new }, P_1} \geq Y_i\end{array}\right.$            (6)

In the range [1, N], k is a randomly chosen natural number, $Y_p$ represents the objective function value associated with the northern goshawk, and P_i denotes the position of the prey for the ith goshawk. During the initial phase of the NGO algorithm, the updated objective function value for the ith candidate solution is denoted by $Y_i^{\text {new, } P_1}$, with the corresponding updated position given by $W_{i, j}^{\text {new, } p_1}$. The parameter r is a random value uniformly distributed in [0, 1], and the index i is randomly selected from the set {1, 2} [29].

Phase 2: Exploitation Operation Chase and Escape: Northern goshawks pursue fleeing prey, skillfully tracking them before attacking. In the NGO-inspired algorithm, this behavior enhances local search precision within a defined attack radius R. Figure 3 shows the pursuit, while Eqs. (6)-(9) model this second phase [29].

$W_{i, j}^{\text {new, } p_2}=W_{i, j}+R(2 r-I) W_{i, j}$             (7)

$R=0.02\left(1-\frac{t}{T}\right)$           (8)

$W_i=\left\{\begin{array}{cc}W_{i, j}^{\text {new }, P_2}, & Y_i^{\text {new }, P_2}<Y_i \\ W_i, & Y_i^{\text {new }, P_2} \geq Y_i\end{array}\right.$                 (9)

The maximum number of iterations is T, and the iteration counter is t. $Y_i^{\text {new, } P_2}$, the goal function’s value, is established by the second step of NGO. The jth dimension is $W_{i, j}^{\text {new, } p_2}$, while the new status for the ith suggested solution is $W_i^{\text {new, } p_2}$.

The proposed NGO algorithm iteratively updates each population member in 50 steps by recalculating the objective function and population values and searching for the optimal solution. Eqs. (4)-(9) are used to update individuals during the iterations until the halting requirement is met. The best outcome of the procedure is then utilized as the quasi-optimal solution to the optimization problem.

3.4.2 Northern goshawk optimization features selection and hyperparameter optimization

The proposed wrapper-based optimization strategy uses the NGO algorithm to choose features and tune base classifier hyperparameters. Each candidate solution is a two-part hybrid vector. The first portion is a binary vector of selected TF-IDF features, with 1 indicating selection and 0 indicating exclusion. The second section includes hyperparameter values for RF, AdaBoost, SVM, and Naive Bayes classifiers. Table 3 lists hyperparameter search ranges and reproducibility-ensuring random seed values. A feature subset from the training data is utilized to train the classifier with the provided hyperparameter settings for each candidate solution. The weighted F1-score on the validation set is adopted as the fitness function. NGO iteratively adjusts the population to maximize classification performance and minimize feature subset dimensionality. The stacking ensemble model uses the best feature subset and hyperparameter configuration for each classifier as base learners after optimization. In addition, Table 4 provides a concise overview of the quantitative attributes of the five datasets. These attributes include the following: total samples, training and testing splits, vocabulary sizes before and after preprocessing, TF-IDF dimensionality, and the number of features preserved after NGO-based feature selection.

3.png

Figure 3. Northern goshawk-prey chase [29]

Table 3. Northern goshawk optimization (NGO) hyperparameter search ranges and random seed values

Classifier	Hyperparameter	Search Range
Random Forest	n_estimators	50 – 200
	max_depth	5 – 30
	random_state	{42, 123, 2024}
AdaBoost	n_estimators	50 – 200
	learning_rate	0.01 – 2.0
	random_state	{42, 123, 2024}
Support Vector Machine	C	0.1 – 10.0
	kernel	{linear, rbf}
Naïve Bayes	alpha	0.001 – 1.0
NGO NGO NGO	Population Size	30
	Maximum Iterations	50
	Fitness Metric	Weighted F1-score
	Random Seeds	{42, 123, 2024}

3.5 Machine learning classification algorithms

The following section describes several classification ML algorithms:

3.5.1 Random Forest classifier

RF is a flexible, simple, and robust supervised machine learning method capable of handling both classification and regression tasks. It works by constructing a collection of decision trees (DTs) that operate collectively to improve prediction accuracy. In classification, each tree independently predicts the class label, and the final output is determined by majority voting among all trees [30].

3.5.2 AdaBoost

AdaBoost is a classification algorithm that combines multiple weak learners, usually DTs, into a strong classifier. Each training instance starts with equal weight, and misclassified instances receive higher weights in subsequent rounds. The weighted errors of weak learners are used to build the final strong classifier, as shown in Eq. (10).

$H(x)=\operatorname{sign}\left(\sum_{t=1}^T \alpha_t \cdot h_t(x)\right)$            (10)

The base model is h_t(x), while the strong model is H(x). Let αt stand for each model’s weights and sign to define. In order to increase classification accuracy, AdaBoost highlights data that weak classifiers have misidentified, and the final total can be either positive or negative (-1, +1). AdaBoost’s weak base classifiers are well-liked for resolving challenging classification problems [31].

3.5.3 Support Vector Machine

SVM is a supervised learning algorithm effective for classification. It finds an optimal hyperplane to separate classes. Using kernel functions, SVM handles linear and non-linear data by mapping inputs to higher-dimensional spaces, enabling complex decision boundaries for classifying new samples [32-34].

3.5.4 Naive Bayes algorithm

Naive Bayes calculates conditional probability, predicting whether one event will occur if another has already occurred. This classification approach employs Bayes’Theorem and predictor independence. In NB, one feature within a class is independent of the others. Naive Bayes is efficient, simple to implement, and effective for large datasets. Text classification using binary and multiclass classifications is accurate [35, 36].

3.6 Proposed SMOTE-NGO stacking ensemble framework

To enhance classification and generalization, advanced ensemble approaches, such as stacking ensemble learning, integrate predictions from base classifiers. To improve the results of the basic model, stacking employs a meta-learner, a second-level classifier, as opposed to bagging and boosting. For this study's base learning, four top-notch machine learning classifiers were used: RF, AB, SVM, and NB. The NGO algorithm chose features and optimized hyperparameters for each base classifier before training the model. To train the improved classifiers, a feature subset from TF-IDF vectorization and SMOTE oversampling was utilized. as in Table 5, which compares the five cyberbullying datasets' class distributions before and after SMOTE.

The table displays the degree of class imbalance in each dataset and the amount of synthetic minority-class samples produced by SMOTE to achieve balanced training data. Cyberbullying incidents were significantly underrepresented in all five datasets prior to oversampling. Dataset 1 had 81,074 training samples; 10,347 were from the minority class and 70,727 were from the majority class. This is to demonstrate the point. To maintain a balanced dataset, SMOTE generated 60,380 synthetic samples, bringing the total number of samples to 141,454. With 10,753 minority class samples and 100,992 majority class samples, Dataset 3 also showed the most severe discrepancy. In order to establish a balanced collection of 201,984 cases, SMOTE generated 90,239 synthetic samples. This procedure was also applied to the other datasets. Dataset 2 added 2,225 hypothetical examples to increase the number of samples from 6,153 to 8,378. The addition of 6,504 synthetic samples raised the sample size of Dataset 4 from 9,262 to 15,766, while the addition of 5,521 synthetic instances raised the sample size of Dataset 5 from 10,245 to 15,766. After applying SMOTE to all datasets, the numbers of occurrences from the majority and minority classes were identical. Training classifiers, tuning hyperparameters, and picking features were all made easier once this balancing phase reduced the model's bias toward the majority class. Apply the NGO algorithm for feature selection and hyperparameter tweaking during training, as I explained in Section 3.4.2. All of the base classifiers make predictions about the training samples. A meta-feature matrix is constructed for the purpose of training meta-learners by merging these outputs. To make a classification decision in testing, the trained meta-learner takes the meta-feature vector from the basic classifier predictions and applies it.

Table 4. Quantitative statistics of preprocessing, TF-IDF, and northern goshawk optimization (NGO)-based feature selection across five datasets

$Z=\left[h_1(x), h_2(x), \ldots \ldots, h_n(x)\right]$                (11)

where, Z stands for the base learners' predictions. The final prediction is then generated by training a meta-classifier g(⋅) on Z, as represented mathematically by Eq. (12).

$\overbrace{y}=g(Z)$               (12)

where, $\overbrace{y}$ denotes the final predicted class label.

The suggested stacking ensemble offers a strong and efficient framework for cyberbullying detection on datasets with different class imbalance ratios by combining the strengths of numerous optimal classifiers.

4.png

Table 8. Analysis result models performance in dataset 1

Table 9. Analysis result models performance in dataset 2

Table 10. Analysis result models performance in dataset 3

Table 11. Analysis result models performance in dataset 4

Table 12. Analysis result models performance in dataset 5

3.9.2 Learning curve analysis

The training behavior and generalizability of the proposed SMOTE-NGO stacking ensemble were examined by creating learning curves as training samples were amassed. Each dataset's convergence behavior and model stability may be observed in the graphs as illustrated in Figures 5–9, which display the training and validation accuracy over larger portions of the training data. The outcomes demonstrate that the NGO algorithm improved the stacking ensemble hyperparameters for better prediction performance and generalization. Training accuracy stayed near 100% across all five datasets, whereas validation accuracy rose steadily as training data increased. Without overfitting or underfitting, the model learns resilient decision constraints, as seen by the slight and declining gap between the two curves. The confidence intervals of the validation curve shrank as the sample size grew, suggesting less variation and more reliable model predictions. More training cases stabilize the ensemble, as this pattern indicates. These findings are validated by an analysis that is specific to the dataset considered. Training on Dataset 1 was nearly flawless, with validation accuracy increasing from 98.0% to 99.7%. Just like in Dataset 2, validation accuracy went up from 96.6% to 99.5%, indicating effective learning even with a lower baseline. Dataset 3's validation accuracy was the highest at almost 99.8 percent, and it nearly overlapped the training curve, suggesting very little variation and excellent generalizability. Validation accuracy increased from 94.5% to 98.5% on the most challenging dataset 4, demonstrating the model's ability to handle increasingly complicated data distributions. Validation accuracy increased from 97.3% to 99.4% as a result of a smaller training-validation gap in Dataset 5, which was achieved by expanding the training set. The learning curves demonstrate that the suggested SMOTE-NGO stacking ensemble performs admirably across different abuse datasets, remains stable with more data, and improves with time. This model is a strong and accurate cyberbullying detector, as shown by its near-saturated training accuracy, steadily increasing validation performance, and shrinking uncertainty bands.

3.9.3 Compare Receiver Operating Characteristic curve analysis

The ROC curves were used at various discrimination thresholds to assess the model's categorization ability. Model discrimination is measured by the AUC, and the trade-off between TPR and FPR is shown by the ROC curve. Positive upper-left curves and AUC values indicate effective cyberbullying detection. Figures 10-14 illustrate datasets (1)-(5). ROC curves and AUC values were generated after optimization. The performance analysis for each dataset is based on the figures provided: The stacking ensemble model with NGO achieved 99.95% AUC and excellent discrimination and performed best for Dataset 1 (DS1). As shown in Figure 10. Compare ROC and area curves for ML and ensemble models after optimizing ds1.

Despite its performance, Naive Bayes had a lower AUC than the stacking ensemble model. AUC was 94.06% for the stacking ensemble model in Dataset 3 before optimization. It was great to get 100% after optimization in the last analysis. As shown in Figure 12. Compare ROC and Area curves for ML and ensemble models after Optimized ds3.

5.png

Figure 5. Learning curve of proposed SMOTE-NGO stacking ensemble ds1

6.png

Figure 6. Learning curve of proposed SMOTE-NGO stacking ensemble ds2

7.png

Figure 7. Learning curve of proposed SMOTE-NGO stacking ensemble ds3

8.png

Figure 8. Learning curve of proposed SMOTE-NGO stacking ensemble ds4

9.png

Figure 9. Learning curve of proposed SMOTE-NGO stacking ensemble ds5

In Dataset 4, Dataset 4’s stacking ensemble model had 96.41% AUC before optimization and 99.90% after optimization. As shown in Figure 13. Compare ROC and Area curves for ML and ensemble models after Optimized ds4.

The entire stacked ensemble model scored 97.63% AUC and achieved the highest performance for Dataset 2 after tuning. All other models achieved an AUC above 89%. As shown in Figure 11. Compare ROC and Area curves for ML and ensemble models after Optimized ds2.

10.png

Figure 10. Compare Receiver Operating Characteristic (ROC) and Area curves for machine learning (ML) and ensemble models after optimized ds1

11.png

Figure 11. Compare Receiver Operating Characteristic (ROC) and area curves for machine learning (ML) and ensemble model after optimized ds2

12.png

Figure 12. Compare Receiver Operating Characteristic (ROC) and area curves for machine learning (ML) and ensemble model after optimized ds3

The stacking ensemble model wins Dataset 5 with 92.07% AUC before optimization and 99.71% after. Naive Bayes is superior in this data type, ranking second. Another second place for Naive Bayes shows stability. Perfect models support class differentiation with a small, tailored dataset. As shown in Figure 14. Compare ROC and Area curves for ML and ensemble models after Optimized ds5.

13.png

Figure 13. Compare Receiver Operating Characteristic (ROC) and area curves for machine learning (ML) and ensemble model after optimized ds4

14.png

Figure 14. Compare Receiver Operating Characteristic (ROC) and area curves for machine learning (ML) and ensemble model after optimized ds5

3.9.4 Fold stratified cross-validation results

On the training data, 5-fold stratified cross-validation was used to assess the generalization performance of the proposed SMOTE-NGO Stacking Ensemble. This method involved dividing the training set into five equal parts, or "folds," and maintaining the default class distribution for each. A validation set was used once for each fold, and the other four folds were utilized for training the model. A total of five evaluations were conducted on the model. When evaluating predictive performance and consistency among folds, the assessment parameters were averaged, and their standard deviations were determined. Tabulated in Table 13 are the aggregated cross-validation findings from the five benchmark datasets.

The model's strong performance across all evaluation measures showed great generalization across data divisions. With an accuracy of 99.88% ± 0.10, a precision of 99.89% ± 0.10, a recall of 99.97% ± 0.03, and an F1-score of 99.93% ± 0.06, Dataset 3 demonstrated the best cross-validation performance. With an accuracy of 99.71% ± 0.03 and an MCC of 97.37% ± 0.72, Dataset 1 produced the best results. In a similar vein, Dataset 2 showed consistent performance with an accuracy of 99.51% ± 0.31 and an MCC of 95.15% ± 0.43 Dataset 4, with an accuracy of 98.46% ± 0.25, an F1-score of 98.48% ± 0.25, and an MCC of 97.94% ± 0.10, fared well despite the difficulties it presented. With an accuracy of 99.38% ± 0.20 percent and an F1-score of 99.67% ± 0.10 percent, Dataset 5 proved to be quite accurate. The small standard deviations of all measures demonstrate the excellent generalizability and stability of the proposed model between folds.

Table 13. Results of 5-fold stratified cross-validation on five different datasets

Metric	Dataset 1 (5-Fold Stratified CV)	Dataset 2 (5-Fold Stratified CV)	Dataset 3 (5-Fold Stratified CV)	Dataset 4 (5-Fold Stratified CV)	Dataset 5 (5-Fold Stratified CV)
Accuracy	99.71% ± 0.03	99.51% ± 0.31	99.88% ± 0.10	98.46% ± 0.25	99.38% ± 0.20
Precision	99.88% ± 0.05	99.63% ± 0.28	99.89% ± 0.10	98.33% ± 0.33	99.62% ± 0.14
Recall	99.82% ± 0.07	99.88% ± 0.11	99.97% ± 0.03	98.64% ± 0.47	99.72% ± 0.08
F1-Score	99.85% ± 0.02	99.75% ± 0.16	99.93% ± 0.06	98.48% ± 0.25	99.67% ± 0.10
ROC-AUC	99.67% ± 0.33	98.43% ± 1.62	99.98% ± 0.02	99.87% ± 0.03	99.70% ± 0.14
PR-AUC	99.98% ± 0.02	99.94% ± 0.07	99.98% ± 0.002	99.84% ± 0.04	99.98% ± 0.01
MCC	97.37% ± 0.72	95.15% ± 0.43	98.31% ± 1.93	97.94% ± 0.10	94.46% ± 1.76

3.9.5 Stability analysis across random seeds

A stability study was conducted to evaluate the predicted model's repeatability when given random initializations. The experimental pipeline that included train-test splitting, SMOTE oversampling, NGO-based optimization, and stacking ensemble training was repeated three times with different seeds: 42, 123, and 2024. The assessment metrics were then normalized to determine the model's resilience and sensitivity to stochastic variation. An extremely low standard deviation indicates consistent and dependable performance throughout all tests. Table 14 provides a numerical summary of the results across all five datasets, while Figure 15 displays the Stability Analysis of Evaluation Metrics Across the Five Datasets.

The suggested SMOTE-NGO stacking ensemble consistently outperformed all assessment criteria, with minimal standard deviations, even when the initialization and data partitioning adjustments were random. Dataset 3 demonstrated the most stable results, with an accuracy of 99.80% ± 0.23%, an F1-score of 99.89% ± 0.12, a ROC-AUC of 99.99% ± 0.02, a PR-AUC of 100% ± 0, and an MCC of 98.31% ± 1.93. The high accuracy (99.51% ± 0.31 for dataset 1 and 99.74% ± 0.03 for dataset 2) and robustness were shown by these datasets. With a recall of 99.16% ± 0.18, an accuracy of 98.97% ± 0.05, precision of 98.81% ± 0.27, and MCC of 97.94% ± 0.10, the model did well on Dataset 4. The reliability was strong in dataset 5, with an accuracy of 99.33% ± 0.20 and an F1-score of 99.64% ± 0.11. Cyberbullying detection reliability is demonstrated by these results, which also demonstrate the robustness and consistency of the suggested methodology across random seeds.

Table 15 compares the datasets, techniques, optimization tactics, and performance measurements of current studies on cyberbullying detection. The SMOTE-NGO stacking ensemble outperforms state-of-the-art methods on a wide variety of benchmark datasets.

15.png

Figure 15. Stability analysis of evaluation metrics across the five datasets

Table 14. Stability analysis results of different seed number across five datasets

Metric	Dataset 1 (Stability Analysis)	Dataset 2 (Stability Analysis)	Dataset 3 (Stability Analysis)	Dataset 4 (Stability Analysis)	Dataset 5 (Stability Analysis)
Accuracy	99.51% ± 0.31	99.74% ± 0.03	99.80% ± 0.23	98.97% ± 0.05	99.33% ± 0.20
Precision	99.63% ± 0.28	99.81% ± 0.07	99.84% ± 0.29	98.81% ± 0.27	99.52% ± 0.33
Recall	99.88% ± 0.11	99.92% ± 0.09	99.95%± 0.06	99.16% ± 0.18	99.77% ± 0.13
F1-Score	99.75% ± 0.16	99.86% ± 0.02	99.89%± 0.12	98.98% ± 0.05	99.64% ± 0.11
ROC-AUC	98.43% ± 1.62	99.05% ± 0.26	99.99% ± 0.02	99.91% ± 0.04	99.60% ± 0.21
PR-AUC	99.94% ± 0.07	99.96% ± 0.02	100% ± 0	99.88% ± 0.06	99.96% ± 0.02
MCC	97.37% ± 0.72	95.66% ± 0.51	98.31% ± 1.93	97.94% ± 0.10	94.46% ± 1.76

Table 15. Comparison of cyberbullying detection studies

No. Ref	The Purpose of the Study	Technique Used	Dataset Type	Finding and Identifying Features	The Best Result
[10]	Lightweight multilingual hate speech detection.	DDML + XLM-R.	9 multilingual hate speech datasets, including Twitter Racism and Sexism datasets.	Contextual transformer features.	Acc. 93.64%, F1 87.95%.
[11]	Detect cyberbullying on Twitter.	HRecRCNN + MFFA.	4,556 Twitter posts.	TF-IDF, BoW, N-grams.	Accuracy 95.0%.
[12]	Detect online hostility.	DNN + Word2Vec.	20,001 English tweets.	Emotional + Word2Vec features.	F1-score 97.0%.
[13]	Bengali cyberbullying detection.	XLM-RoBERTa.	2,751 Bengali social media texts.	TF-IDF, BoW, embeddings.	Acc. 82.61%, F1 0.83.
[14]	Detect toxic Hinglish messages.	RF, SVM, KNN, AdaBoost.	18,307 Hinglish tweets.	TF-IDF features.	RF F1 97.2%.
[15]	Compare RoBERTa with ML and DL models.	Fine-tuned RoBERTa.	48,000 annotated tweets.	Contextual transformer features.	Accuracy 83.9%.
[16]	Detect cyberbullying from images.	OCR + TF-IDF + ML classifiers.	160,000 samples from YouTube, Twitter, Wikipedia, and Kaggle.	OCR-extracted text + TF-IDF.	Accuracy 96.0%.
[17]	Compare BERT and CNN-SVM models.	BERT, CNN-SVM.	Two benchmark datasets.	Contextual and CNN-based features.	BERT Accuracy 97.34%.
[18]	Develop a real-time explainable detection framework.	ARFC + GPT-4o-mini.	15,890 English tweets from X.	LLM-generated semantic + sentiment features.	F1-score 90.06%.
[19]	Improve cyberbullying detection using aggression-enhanced prompts.	EPP + LoRA + Gemma.	Five aggression datasets + 47,000 tweets.	Aggression labels as contextual cues.	Macro-F1 0.99.
Proposed Study	Develop an optimized cyberbullying detection framework.	SMOTE-ENN + NGO + Stacking Ensemble.	Five benchmark datasets (Twitter, Kaggle, Wikipedia).	TF-IDF with optimized feature selection and hyperparameter tuning.	Proposed Study achieved over 5-fold stratified cross-validation produced an F1-score of 99.93% ± 0.06 and an accuracy of 99.88% ± 0.10%, AUC 100%.