The vehicle routing problem (VRP), especially its capacitated (CVRP), has been extensively studied. In the CVRP, each customer must be served exactly once without exceeding the vehicle’s capacity, with the aim of minimising the total distance of all routes. As a nondeterministic polynomial (NP)-hard problem, the CVRP is best solved using metaheuristics. The red deer algorithm (RDA), inspired by red deer mating behaviour, is population-based search metaheuristic with both exploration and exploitation phases. However, it may still get trapped in local optima and lacks a strong intensification mechanism for refining solutions. In RDA, not all solutions are explored, as it imposes no limit on solution reproduction, thereby restricting exploration. Thus, this study proposes an enhanced RDA that incorporates adaptive memory strategies to strengthen exploration and exploitation. The RDA classifies the population into males (i.e., high-quality solutions) and females or hinds (i.e., lower-quality solutions), with males further divided into commanders and stags. Exploitation occurs through roaring, selecting commanders, and fighting, while exploration involves forming harems and mating commanders with hinds both within and outside their harems. In enhanced RDA, previously mated hinds are adaptively excluded to promote mating with new ones, thereby enhancing exploration and solution quality. Enhanced RDA with an adaptive memory strategy (RDAM) was tested on 87 benchmark instances, outperforming a construction algorithm by 96.5% and the original RDA by 87.3%, and showing competitive results with established methods. The adaptive memory strategy in RDA enhanced the total distance compared with the original RDA, while improving its exploration–exploitation in solving the CVRP, offering theoretical and practical benefits for logistics efficiency.
Real-time object tracking in embedded systems requires algorithms that balance processing performance with energy efficiency for practical deployment. This study evaluates the performance of the TG-Lite algorithm on Jetson Nano embedded hardware for surveillance applications requiring a minimum processing rate of 10 frames per second (FPS). The methodology involves comprehensive performance analysis, measuring frame-processing rates, energy consumption patterns, and system resource utilisation across video-file and live-stream scenarios. TG-Lite achieves 15.51 FPS processing capability with 11.24 watts (W) power consumption, meeting real-time processing requirements for embedded surveillance deployment. The algorithm demonstrates consistent performance across different input conditions, maintaining stable processing rates on both video files (10.86 FPS) and live streams (15.51 FPS). TG-Lite exhibits efficient resource utilisation, with moderate central processing unit (CPU) usage (17.4% average) and effective graphics processing unit (GPU) acceleration. The energy analysis reveals TG-Lite's 1.38 frames-per-watt efficiency, indicating effective allocation of computational resources for real-time processing demands. Performance evaluation shows that TG-Lite maintains stable tracking continuity, with an average track length of 35.8 frames and a consistent confidence distribution pattern. These results demonstrate the suitability of the proposed approach for embedded computer vision applications under the evaluated edge constraints.
Fuzzy cooperative games using Type-1 fuzzy numbers are commonly applied to model vagueness and imprecision in coalition values and player allocations. However, such models are less effective in capturing increased imprecision. This paper introduces interval Type-2 Gaussian fuzzy numbers (IT2 GFNs) to better represent increased imprecision in coalition values and player allocations in distributed energy systems (DES) caused by uncoordinated electric vehicle (EV) charging and discharging patterns. An analytical solution is proposed to ensure fair and efficient player allocations in IT2 Gaussian fuzzy cooperative games. The proposed method avoids the computational challenges of IT2 subtractions and the associated Type-1 fuzzy number subtractions, which tend to amplify the footprint of uncertainty (FOU) or restrict applicability. A comparative analysis with the interval-based Shapley value, computed using Moore's subtraction operator, demonstrates that the proposed method achieves allocation efficiency by fully distributing the grand coalition value among players while overcoming the known limitations of the Hukuhara difference, which is not always well-defined.
Rice farming faces persistent challenges that constrain productivity and sustainability, underscoring the urgent need for digital innovations such as the Internet of Things (IoT). While IoT technologies offer the potential for efficiency and resilience, their adoption depends on farmers' willingness to pay (WTP), which remains poorly understood in resource-constrained contexts. This study aimed to identify the determinants of WTP for IoT adoption by integrating farmer-related, technological, and institutional factors. A structured survey was conducted among 229 rice farmers in the Integrated Agricultural Development Area, Barat Laut Selangor, Malaysia. The findings revealed that most farmers demonstrated moderate IoT awareness, with the awareness index highlighting gaps in technical knowledge and training. The IoT readiness index recorded relatively high scores for shared infrastructure (71.8) and Internet connectivity (71.6) but lower levels for technical support (60.0) and energy supply (54.4), reflecting uneven preparedness. Logistic regression results indicated that higher-income farmers were significantly more likely to invest, with middle-and high-income groups showing 20.13% and 22.45% higher probabilities, respectively. Secure land tenure increased WTP by 3.18%, while each additional year of age reduced WTP by 0.61%. IoT infrastructural readiness raised WTP by 10.53%, and perceived ease of use increased WTP by 8.27%. However, perceived usefulness exhibited a negative marginal effect, suggesting that recognising IoT's benefits does not necessarily translate into WTP when financial or institutional barriers persist. The findings indicate that financial support, reliable IoT infrastructure, and practical training are crucial to fostering inclusive IoT adoption and enhancing resilience in rice farming.
The increasing prevalence of online gaming among adolescents has heightened the need to understand the psychological factors that influence secure online behaviour. Protection Motivation Theory (PMT) is limited in its ability to explain technology-related behaviours because it primarily focuses on conscious, rational decision-making driven by threat and coping appraisals. While PMT is widely used to explain individuals' protective intentions, emerging cognitive constructs such as cybercognition, which refers to the skills, heuristics, and cognitive dispositions that shape online decision-making, remain underexplored within this framework. This study investigates cybercognition as a potential determinant of the secure intention behaviour model, which refers to self-instructions to perform actions to attain secure behaviour in an online gaming environment. Using a simple random survey of adolescent gamers, the research examines how cybercognitive abilities interact with PMT components, including perceived severity, vulnerability, self-efficacy and response efficacy, to predict intentions to engage in secure online practices. Structural equation modelling reveals that cybercognition significantly enhances the explanatory power of PMT by influencing both threat appraisal and coping appraisal pathways. Findings suggest that adolescents with higher cybercognitive awareness are more likely to adopt protective behaviours while gaming. The study contributes to theoretical advancements in PMT by integrating cybercognition and offers practical implications for designing targeted digital safety interventions and educational programs for the youth.
Recommender systems play a pivotal role in enhancing user experience and driving engagement across e-commerce platforms. However, traditional approaches such as collaborative filtering and matrix factorisation often struggle with sparse user-item interactions and fail to capture contextual nuances. To address these limitations, this paper proposes a novel hybrid framework that integrates Graph Neural Networks (GNNs) and Transformer-based architectures for context-aware recommendation. The framework first constructs a dynamic bipartite graph from user interactions and product metadata, enabling the GNN to learn relational embeddings that reflect user preferences and item similarities. These embeddings are then fused with a Transformer module that applies multi-head self-attention to model temporal and semantic patterns within user behaviour sequences. We propose a hybrid Graph-Augmented Transformer framework for context-aware recommendation, evaluated on the Amazon Reviews 2023 dataset (571M reviews, 54.5M users, 48.2M items). The model integrates graph-based relational embeddings with transformer-based sequential attention. Compared to LightGCN, SASRec, and BERT4Rec baselines, our approach achieves significant improvements in NDCG, recall, and robustness under cold-start conditions. The framework is scalable and suitable for real-time deployment, offering a practical blueprint for next-generation recommender systems. By combining graph-based relational learning with attention-driven context modelling, this research contributes a powerful and flexible solution for next-generation recommender systems.
Heat-related mortality has emerged as a critical public health issue, driven by the accelerating impacts of climate change. Most studies establish a direct causative relationship between high temperatures and mortality; however, there is scarce literature studying the climate change mortality ellipse. This study aims to fill this gap by examining the effects of climate change on mortality in the tropical region with consistently high year-round temperatures, specifically in Malaysia. Structural time-series models were applied to annual mortality data for the period 2005 to 2022, obtained from the Department of Statistics Malaysia. Monthly climate data, including temperature, rainfall amount, number of rainy days, and relative humidity, were sourced from the Malaysian Meteorological Department and subsequently aggregated into annual values to ensure consistency in the analysis. The model uses impulse indicator saturation, which makes it easier to spot structural breaks and extreme values, improving the reliability of the results. The analysis indicates that higher rainfall is strongly associated with increased mortality, reflecting the health risks linked to flooding and waterborne diseases. In contrast, periods of higher relative humidity tend to correspond with lower mortality rates. The fully saturated model identifies two key structural shifts in 2006 and 2018, likely caused by abrupt changes in temperature and rainfall. These findings provide a solid foundation for targeted interventions, such as heat-stress regulations and localised air-quality measures, and offer evidence to guide strategies to reduce climate-related health risks in the country, while also supporting broader public health planning.
Financial statement fraud detection is critical to maintaining trust among investors, regulators, and analysts. However, traditional audit procedures often fail to detect anomalies effectively because they occur infrequently but can result in significant economic losses. This study proposes an oversampling approach using a modified threshold in the Wasserstein Generative Adversarial Network with Gradient Penalty (WGANGP) to enhance synthetic data variance in financial fraud detection. The financial data were collected from the financial reports of companies listed on the Indonesia Stock Exchange and were labelled according to the Balanced Scorecard framework into four categories: normal, alarm, risky, and fraud. Given the severe class imbalance, this study introduces a WGANGP model with threshold optimisation in the generator and a gradient penalty to generate high-quality synthetic samples. This study conducted general and per-entity oversampling scenarios and evaluated them using Euclidean distance, Wasserstein distance, and classification metrics. In Scenario 1, the Generative Adversarial Network (GAN) outperformed the Synthetic Minority Oversampling Technique (SMOTE) and vice versa in Scenario 2. However, in financial fraud detection, the WGANGP with enhanced thresholding improved the F1-score by 13% to 17% compared to SMOTE and five GAN-based models across thirteen classification models, including traditional, machine learning, and deep learning models. This finding suggests that optimising the threshold in WGANGP reduces variance and improves model performance. Furthermore, generating synthetic data that is very similar to actual data may not necessarily improve classification; therefore, it is necessary to test how oversampling affects subsequent stages.
Effective self-care practices are important for university students' well-being. However, many students struggle to sustain these practices due to low motivation with existing digital self-care tools. While gamification shows potential, its effectiveness is often limited by implementations that lack robust theoretical foundations or alignment with psychological and motivational theories. This study aims to utilise gamification to encourage university students to practice self-care. The study evaluates two versions of a self-care application prototype: Version A (without gamification) and Version B (with gamification), utilising the Wheel of Sukr framework. A pilot study was conducted to refine experimental procedures, followed by a main experiment with 20 undergraduate students. Data collection included pre-test forms, the User Engagement Scale-Short Form (UES-SF), the System Usability Scale (SUS), and qualitative feedback. Wilcoxon Signed-Rank tests showed significant differences favouring Version B (Z =-2.354, p = 0.019) for user engagement metrics, including focused attention, aesthetic appeal, and reward factors. Qualitative feedback indicates that students found gamification elements such as points, badges, and weather-based recommendations engaging, though some noted the added complexity as a usability challenge. The average SUS score for Version B was 80.63, indicating high usability. The findings suggest that adapting the Wheel of Sukr framework may provide a structured design approach for integrating gamification into student self-care applications, enhancing engagement while maintaining usability. This approach offers practical implications for the design of future theory-driven digital health interventions.
Accessing extensive and varied datasets is essential for developing strong predictive models in data analytics. However, many real-world applications suffer from small and imbalanced datasets, leading to overfitting, poor generalisation, and low model performance. Traditional data augmentation techniques are often unsuitable for tabular data, as they fail to preserve complex feature relationships. To address this challenge, this study adapts the Conditional Tabular Generative Adversarial Network (CTGAN) for synthetic data generation. The proposed approach involves five phases: (1) Data Acquisition, 2) Data Preparation, (3) Model Training, (4) Synthetic Data Generation, and (5) Evaluation. Experimental results on three benchmark datasets show that the proposed work produced data that closely adheres to the statistical distribution of the original dataset, with Wasserstein Distance < 0.05 for numerical features and Jensen-Shannon Divergence < 0.08 for categorical features. Additionally, models trained on datasets including synthetic and real data achieved up to 15% improvement in classification accuracy compared to those trained on real and small datasets alone. Training on a combination of real and synthetic data for the minority class in large datasets significantly improves the F1-score, with gains of approximately 9–10%. This approach also yields a modest increase in overall accuracy (around 1.5%), suggesting enhanced model generalisation. These results indicate that the adapted CTGAN is a viable option for data augmentation, addressing problems with limited and imbalanced data for machine learning data training.
Digital advertising continues to grow rapidly, yet advertisers face persistent uncertainty in selecting the most effective ad format for different campaign objectives and audience segments. Prior studies often rely on limited metrics or lack interpretability, making it difficult to explain why certain formats perform better. This study addresses this gap by evaluating ad format effectiveness using explainable machine learning. This study evaluates the effectiveness of image and video advertisements (ads) across five key performance indicators (KPIs): reach, impressions, link clicks, cost per click (CPC), and cost per thousand impressions (CPM). A dataset of 4,526 campaigns from Meta Ads Manager (July 2021–August 2024) was analysed using machine learning models integrated with Shapley Additive Explanations (SHAP). Model performance was assessed using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R-squared (R²). The results showed that video ads achieved higher reach and impressions with lower CPM. Meanwhile, image ads delivered lower CPCs and stronger click performance, particularly among users aged 13–17 and 55+. Aside from confirming established format patterns, this study made certain contributions by applying explainable machine learning in a large-scale, non-Western context and by clarifying the mechanisms behind ad performance. The findings offered actionable guidance; for example, video ads were optimal for awareness and visibility objectives. Additionally, image ads were preferable for engagement-driven campaigns.
The increasing sophistication of fraudulent activities in digital financial systems necessitates real-time anomaly detection models that can adapt to evolving transactional behaviours. While prior research has explored machine learning approaches for fraud detection, most rely on static datasets and overlook temporal dependencies inherent in financial transaction streams. This study addresses this critical gap by proposing a hybrid anomaly detection model that integrates dynamic streaming-frequency analysis via a 7-day sliding window, an unsupervised Long Short-Term Memory (LSTM)-autoencoder for anomaly scoring, and a supervised Artificial Neural Network (ANN) for classification. The sliding window mechanism enables the model to capture short-term temporal fluctuations and behavioural patterns, aligning with the streaming nature of financial data. The LSTM-autoencoder is trained exclusively on normal transaction sequences to learn temporal dependencies and compute reconstruction errors, which serve as deep anomaly features. These features are then fed into the ANN to classify transactions as normal or anomalous. Experimental results on the IBM Anti-Money Laundering (AML) dataset demonstrate the effectiveness of the proposed framework, achieving a classification accuracy of 99.92%, precision of 94.12%, recall of 88.43%, and F1 score of 91.19%. This layered architecture not only enhances early detection of anomalous behaviour but also provides a scalable, adaptive solution for real-time fraud detection in streaming financial environments.
In emergency information systems, expert decision-making is a critical process that often employs linguistic terms to convey subjective judgments. Membership functions (MFs) are an essential tool for representing the meaning of these terms, which allows them to be processed computationally. Existing methods for eliciting MFs of these terms struggle to capture the inherent variability and probabilistic nature of expert opinions, hindering accurate representation of uncertainty. This study proposes a framework for eliciting MFs of probabilistic linguistic terms using the Interval Estimation (IE) technique. The procedure integrates a graphic survey to construct MFs that fulfil the definition of Triangular Fuzzy Number (TFN). This approach enables experts to directly express their uncertainty ranges, resulting in a high level of consistency and thereby validating the shared and context-dependent understanding. A case study on eliciting MFs and integrating them with a Bayesian Network (BN) decision model for emergency evacuation, demonstrating the utility of our proposed framework. The results indicate that an evacuee's ability to perceive fire cues and hazardous events, while maintaining psychological stability, increases the probability of a successful evacuation. The performance analysis, as indicated by sensitivity values, confirms the stability and robustness of the BN model parameters, thereby validating the rationality and meaningfulness of the MFs. The resulting MFs demonstrated a significant improvement in capturing the uncertainty of expert assessment in the emergency information system compared to traditional methods.
Decision-making in complex, uncertain environments, like emergency evacuations, often relies on the use of linguistic terms to express subjective judgments. However, the inherent ambiguity of these linguistic terms, coupled with the variability of expert opinions, poses a significant challenge for accurate decision-making. Membership functions (MFs) are essential tools for quantifying and representing the meaning of these linguistic terms, allowing for the computational processing of subjective judgments. Existing methods for eliciting MFs of these terms struggle to capture the inherent variability and probabilistic nature of expert opinions, hindering the accurate representation of uncertainty. Existing methods for eliciting membership functions (MFs) of these terms struggle to capture the inherent variability and probabilistic nature of expert opinions, hindering accurate representation of uncertainty. This study propose a framework for eliciting MFs of probabilistic linguistic terms using the Interval Estimation (IE) technique. The procedure integrates a graphic survey to construct MFs that fulfils the definition of Triangular Fuzzy Number (TFN). This approach allows experts to directly express their uncertainty ranges, ensuring the resulting MFs reflect both their individual perceptions and the overall probabilistic distribution of opinions. A case study eliciting psychological responses in fire evacuation scenarios is utilized to demonstrate the utility of our proposed framework. The developed MFs were integrated into a Bayesian Network (BN) decision model. The performance analysis indicated by sensitivity values confirms the stability and robustness of the BN model parameters, thereby validating the rationality and meaningfulness of the expert-elicited MFs. The resulting MFs demonstrated a significant improvement in capturing the variability of psychological responses compared to traditional methods. This robust methodology provides a practical tool for developing expert-driven fuzzy linguistic scales tailored to specific domains thus offering practical applications for decision-making in uncertain environments.
The growing volume and sophistication of phishing emails have become a significant threat to data security, often serving as the initial vector for data breaches. Most past studies have focused on comparing machine learning models to determine the best-performing algorithm. They often neglect the role of pre-processing, which also contributes to the effectiveness of these models. To address this gap, this study investigates the impact of pre-processing techniques on phishing email detection, aiming to strengthen data protection. Three supervised machine learning algorithms, which are Support Vector Machine (SVM), Random Forest and Decision Tree, were selected to undergo two experimental iterations: one with basic pre-processing and the other with an enhanced pre-processing technique including Synthetic Minority Oversampling Technique (SMOTE), Term Frequency-Inverse Document Frequency (TF-IDF), Singular Value Decomposition (SVD) and cross-validation. Using a dataset comprising 28,747 labelled emails, the models were trained, tested, and evaluated based on accuracy, precision, recall, and F1-score, with further insight gained through confusion matrix analysis. Among the models, Random Forest demonstrated the strongest consistent performance across all metrics, while Decision Tree showed the most notable improvement. Although SVM maintains high recall and precision, it is less responsive to the applied pre-processing techniques. This result demonstrates that pre-processing techniques significantly contribute to the performance of the detection models. Overall, these findings highlight the critical role of pre-processing in enhancing phishing email detection, which contributes to stronger organisational resilience.
E-learning has become a key component of modern education that provides access to digital learning resources. However, the overwhelming volume of content can make it difficult for learners to find materials suited to their needs. This has led to a growing demand for adaptive learning, which personalises content based on learner characteristics. To support this, e-learning platforms adopt recommender systems through machine learning techniques. While effective, these systems often depend heavily on historical data, such as user ratings and interactions, to generate meaningful recommendations. This dependency introduces a significant challenge known as data sparsity, where insufficient interaction data limits the model's ability to provide accurate recommendations. This study addresses this challenge by proposing a hybrid learning resource recommender model that combines collaborative and content-based filtering and introduces the Learning Object Rating Algorithm (LORA). This hybrid approach reduces reliance on user-generated ratings by allowing LORA to generate initial ratings based on the learners' profiles and resource characteristics, thus filling gaps in interaction history. The model was evaluated through experiments assessing its prediction accuracy and relevance of recommendations by using Mean Absolute Error (MAE), Precision, and Recall. Additionally, the performance of the proposed hybrid model was compared with existing hybrid models through a comparative analysis. Results revealed that the proposed model outperformed previous hybrid recommender models, generating better prediction accuracy and recommendation relevance. The integration of a hybrid approach and LORA enabled the model to generate ratings based on learning styles and resource characteristics, mitigating the data sparsity issue and reducing dependence on user-generated ratings.
Garbage management has become an urgent global challenge due to the expected 70% increase in volume between 2025 and 2050, driven by rapid urbanisation and population growth. Low public understanding of garbage sorting and source-based disposal is reflected in Indonesia's inefficient garbage processing. In addition to being ineffective, traditional methods like landfill disposal and incineration present serious environmental hazards. These traditional methods are often ineffective due to resource constraints and time-consuming nature, and cannot be relied upon for extended periods because they damage the environment during the process. To increase the effectiveness and precision of garbage management, this research proposes a garbage detection and classification model using YOLOv11. This research includes data collection and pre-processing, model training, and performance evaluation using metrics such as mean average precision (mAP). This research uses the Trashnet Garbage Classification Dataset, which has 2,524 total images that are divided into six categories. The key technical contributions of this research are to apply additional techniques, such as data augmentation strategies, to the dataset, enabling a comparison between the original and more advanced datasets. The purpose of the data augmentation technique is to improve model generalisation. The results of the evaluation metrics show that the model using the augmented dataset has slightly better performance with an mAP50 value of 97.8% than the model using the original dataset. This model is capable of identifying and classifying accurately all of the categories in the dataset.
Although the Delone and McLean IS Success Model (D&M) can explain the phenomenon of e-participation success (EPS), the model was initially created in an e-commerce setting and thus neglects external factors related to e-government services. To address the gap, this study revisits the D&M by extending it with four socio-cultural constructs of Trust (TR), Anonymity (AN), Nationalism (NT), and Culture (CR). Based on 428 survey data from Malaysian citizens, a hybrid methodology was employed, integrating Partial Least Squares-Structural Equation Modelling (PLS-SEM) and Multi-layer Perceptron (MLP) to capture non-linear relationships and enhance predictive accuracy. While hybrid modelling is common, past studies have often applied limited classification metrics, and the infrequent use of comprehensive metrics, such as Area Under the Receiver Operating Characteristic Curve (AUC-ROC), may potentially affect the reliability and generalizability of these models. A comparative R² analysis between the baseline and enhanced models in this study revealed significant improvements, with R² for e-participation intention (EPI) increased from 0.620 to 0.728, EPS from 0.330 to 0.345, while User Satisfaction (US) remained strong at 0.765. The analysis predicts a 94.80% success rate for e-participation, with the MLP model further demonstrating robust classification performance, achieving an accuracy of 90.1%, a precision of 0.909, a recall of 0.948, an F1 score of 0.928, and an AUC-ROC of 0.955, outperforming other benchmark classifiers. This study contributes theoretically by introducing underexplored socio-cultural variables into the D&M while methodologically extending the hybrid PLS-SEM and MLP through a robust model validation using AUC-ROC.
The rise of online courses, accelerated by the COVID-19 pandemic, has underscored the need for effective educational models capable of addressing the challenges posed by remote learning. This study focuses on the development of sentiment classifiers using the Coursera reviews dataset to evaluate the polarity of student feedback. This research improved student engagement and support in online education by applying sophisticated sentiment analysis techniques. We explored a comprehensive methodology encompassing various pre-processing techniques, advanced tokenisation methods, and a range of deep learning architectures, including Feedforward Neural Networks (FNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), and Bidirectional Encoder Representations from Transformers (BERT)-based models. Each model's performance is optimised through meticulous hyperparameter tuning using the Optuna framework. Results indicated that BERT is the best model, achieving a recall of 97.50% and an accuracy of 96.83%, while Bidirectional LSTM (BiLSTM) closely followed with a recall of 96.55% and an accuracy of 96.71%. In contrast, simpler models like FNN and RNN exhibited lower accuracy (92.83% and 87.83%, respectively). These findings underscore the importance of advanced models in capturing contextual meanings and highlight the effectiveness of leveraging embeddings, attention mechanisms, and tailored pre-processing strategies, which significantly improve sentiment classification performance.
The widespread adoption of digital technology has also resulted in a concerning trend: the rise in online threats targeted at children. As online threats to children continue to increase, parents need to take proactive steps to protect their children's safety on the Internet. This research explores how parents respond to online threats affecting their children by examining their cyber-parenting approaches and coping strategies through Protection Motivation Theory (PMT). This research involved semi-structured interviews with nine parents whose children had experienced online threats. The study revealed that authoritative parenting, which balances autonomy with guidance, was the most adopted approach by parents. Conversely, parents with a high level of digital literacy tended towards an authoritarian style, characterised by rigorous monitoring of their adolescents' online activities. Following online threat incidents, parents adopted both problem-focused coping and emotion-focused coping. These responses reflected varying degrees of threat appraisal and coping appraisal, consistent with PMT constructs. By identifying prevalent parenting styles and coping strategies, this research contributes to a deeper understanding of how parental behaviours influence children's online behaviours and resilience to cyber incidents. The study also found that effective parental mediation not only reduces children's exposure to harm but also plays a critical role in fostering their cybersecurity awareness, empowering them to recognise risks, apply protective behaviours, and navigate digital environments with greater confidence. Such insights are critical for designing effective interventions and educational initiatives aimed at fostering safer digital practices among families, particularly in raising children's awareness.
