This research presents a comparative study of machine learning and deep learning techniques for detecting malicious URLs, a key concern in cybersecurity. While machine learning methods are widely used, limited work has explored the effectiveness of stacked models combining multiple algorithms. This study evaluates the performance of three machine learning models—Random Forest, XGBoost, and LightGBM—and three deep learning models—LSTM, BiLSTM, and GRU. Results show that machine learning models achieved accuracy scores of 91%, 92%, and 92%, respectively, outperforming the deep learning models, which scored 89%, 91%, and 92%. A stacked model combining these approaches further improved accuracy to 93%. These findings underscore the efficiency of machine learning models in malicious URL detection and demonstrate that stacking can enhance precision. The proposed approach offers a promising solution for businesses and individuals aiming to strengthen online security and combat cyber threats more effectively.

Smishing (SMS phishing) is a form of social engineering attack in which adversaries exploit mobile messaging platforms to deceive users into disclosing sensitive information, installing malicious content, or initiating fraudulent transactions. Unlike traditional cyberattacks that target software vulnerabilities, smishing primarily exploits human trust, making it difficult to detect using conventional rule-based or signature-driven security mechanisms. With the rapid growth of mobile communication and digital financial services, smishing has emerged as a significant cybersecurity threat affecting individuals, enterprises, and communication infrastructures. To address this challenge, this paper proposes a cybersecurity-oriented hybrid fusion framework for smishing detection that integrates advanced feature engineering with deep learning and transformer-based language models. The proposed approach combines semantic embeddings from Sentence-BERT, statistical TF–IDF representations, and SMS-specific security features capturing URL behavior, phone number patterns, sentiment manipulation, linguistic anomalies, and named entity indicators commonly associated with phishing attacks. These heterogeneous features are fused through a multilayer feed-forward neural network, enabling the detection of both known and previously unseen smishing strategies. The framework is evaluated on six publicly available SMS datasets containing English and Bangla messages, ranging from 2,772 to 16,572 samples after preprocessing and class balancing. Experiments follow a stratified train–validation–test split (70%–15%–15%) with 5-fold cross-validation and Optuna-based hyperparameter optimization. The proposed model achieves a peak validation accuracy of 99.67%, an average cross-validation accuracy of 99.38%, and a test accuracy of up to 96.54%, with ROC–AUC values exceeding 98%. The results demonstrate that the proposed approach outperforms traditional machine learning, deep learning, and standalone transformer-based models. To enhance model transparency, Local Interpretable Model-agnostic Explanations (LIME) are incorporated to provide interpretable insights into feature-level decision behavior, demonstrating the effectiveness of the proposed framework for real-world smishing detection in multilingual mobile communication environments.

The rapid transition to online learning has intensified the need to understand the multifaceted drivers of student satisfaction in virtual education environments. Student satisfaction is a cornerstone of effective digital education, influencing academic performance, learner engagement, and institutional decisionmaking. As a critical predictor of retention and perceived instructional quality, it directly impacts the sustainability of online learning systems. However, existing studies often lack integrated frameworks that combine subjective expert judgment, global and local model interpretability, and subgroup-specific analysis. To address this gap, this study aims to develop a hybrid decision-analytic and explainable machine learning framework to systematically assess and interpret student satisfaction in virtual classrooms. The proposed framework adopts a layered analytical design in which statistical feature relevance (ANOVA), model-based interpretability (SHAP and LIME), and expert-driven weighting under uncertainty (FAHP) are treated as complementary but non-overlapping components, with TOPSIS used exclusively for decisionlevel satisfaction ranking. A structured survey has been conducted among 1,469 university students, capturing responses across 21 features encompassing demographic attributes, digital readiness, behavioral patterns, and perceived instructional quality. Feature selection has been performed using the ANOVA F-test, while SHapley Additive Explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) have been applied to derive global and local feature importance. To incorporate subjective expert judgment, feature weights have been calculated using the Fuzzy Analytic Hierarchy Process (FAHP), and composite satisfaction scores have been computed using the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS). These scores have been used to segment students via Agglomerative Hierarchical Clustering into low (n = 852), medium (n = 484), and high (n = 133) satisfaction cohorts. Clusterspecific SHAP and LIME analyses have uncovered distinct patterns of feature influence across subgroups. The identified satisfaction drivers align closely with constructs from established models of technology acceptance and learning motivation, including the Technology Acceptance Model (TAM), the Unified Theory of Acceptance and Use of Technology (UTAUT), and Self-Determination Theory (SDT). High satisfaction has been linked to emotional positivity, perceived equivalence to offline instruction, and academic self-efficacy, whereas low satisfaction has been driven by affordability barriers, psychological stress, and misalignment with preferred learning modalities. The proposed framework offers a comprehensive, interpretable, and rank-aware approach to modeling student satisfaction, equipping educational stakeholders with actionable insights for designing more inclusive, equitable, and learner-centered online education environments.

Log parsing is a fundamental prerequisite for automated system monitoring, anomaly detection, and root cause analysis in large-scale distributed environments. However, existing parsing approaches often rely on heuristic rules, manually engineered features, or fixed similarity thresholds, limiting their adaptability to heterogeneous and evolving log structures. To address these challenges, this study presents SemBERT, a fully unsupervised log parsing framework that integrates semantic BERT embeddings, Incremental Principal Component Analysis (IPCA), HDBSCAN clustering, and adaptive centroid-based cluster merging for robust template mining. Unlike conventional methods that employ fixed merging criteria, SemBERT adaptively determines semantic merging thresholds according to the distribution of cluster centroids, reducing parameter sensitivity across datasets. Experiments on the LogHub HDFS benchmark demonstrated that SemBERT achieved perfect Grouping Accuracy (GA = 100.00%), Parsing Accuracy (PA = 95.82 ± 0.44%), and Template Accuracy (TA = 76.59 ± 4.49%), outperforming several classical parsers in semantic template discovery while maintaining competitive parsing performance. Repeated reproducibility experiments using three different random seed initializations (42, 123, and 999) produced identical adaptive thresholds and final template counts across all evaluated data scales, confirming the deterministic behavior and reproducibility of the proposed framework. Additional scalability experiments on randomly sampled subsets ranging from 50K to 1M log entries demonstrated stable threshold estimation and controlled template growth, highlighting the effectiveness of the adaptive semantic merging strategy. To validate the practical utility of the extracted templates, downstream anomaly detection experiments were conducted using representative machine learning, deep learning, and transformer-based models. The semantic templates generated by SemBERT provide informative representations suitable for downstream log analytics. Furthermore, experiments on the BGL dataset indicate that the framework generalizes effectively to heterogeneous logging environments through lightweight dataset-specific preprocessing. These findings suggest that SemBERT provides a scalable and semantically informed solution for unsupervised log parsing in modern distributed systems.

Background: Thalassemia has been recognized as a critical public health issue in Bangladesh, especially among pregnant women, due to its hereditary nature and the lack of early screening infrastructure. Early identification of at-risk individuals is essential to prevent the transmission of this genetic disorder to future generations and to reduce the burden on an already strained healthcare system. Methods: In this study, an innovative framework for thalassemia risk assessment has been developed by integrating Multi-Criteria Decision-Making (MCDM) methods—specifically AHP-TOPSIS—with machine learning algorithms including Random Forest, XGBoost, and CatBoost. Explainable Artificial Intelligence (XAI) techniques such as SHAP and LIME have also been incorporated to improve model transparency and trustworthiness. Real-world clinical and demographic data, consisting of 16 features and 1200 samples, have been collected through a structured survey and processed using rigorous feature selection and ranking methods. Risk stratification has been performed to classify patients into high, medium, and low categories, enabling targeted intervention. Results: Among all models, the XGBoost classifier trained on AHP–TOPSIS–prioritized features achieved a consistent accuracy of 99.28% under stratified 20-fold cross-validation, demonstrating robust diagnostic classification performance. The model predominantly captures hematologic patterns characteristic of thalassemia manifestations, functioning as an assistive diagnostic framework rather than a causal risk predictor. The explainability of predictions, ensured through comprehensive visual and statistical analyses, further enhances the model’s clinical transparency and reliability. Conclusions: The proposed MCDM–machine learning framework demonstrates strong potential for improving thalassemia risk assessment, enabling early detection and informed decision-making in maternal healthcare. The proposed framework should be regarded as a preliminary proof-of-concept system that demonstrates the feasibility of integrating Multi-Criteria Decision-Making (AHP–TOPSIS) with advanced machine learning and explainable-AI techniques for thalassemia assessment. Although the model achieved strong diagnostic performance under nested cross-validation, additional external validation and inclusion of causal predictors are required before clinical deployment.

Image inpainting aims to restore missing structures and textures given a corrupted image and mask. Despite strong progress, many methods still struggle to model long-range dependencies while preserving local detail, leading to artifacts or over-smooth results. We present SAMRBNET, which treats inpainting as latent context retrieval and fusion. An encoder–decoder backbone is augmented with (i) self-attention to capture global context, (ii) residual blocks to stabilize training and preserve local structure, and (iii) a lightweight gated fusion in latent space, Fgated = Fres ⊙ Fatt, followed by a dilated context block to widen the receptive field. Experiments on CelebA-HQ, Paris Street View, and ImageNet, SAMRBNET achieves state-of-the-art results on PSNR/SSIM/L1; on Places2 it is competitive—with slightly lower SSIM/L1 than BrushNet (see Sec. IV-D1 and Table III). We further report LPIPS (AlexNet), mean ± std over three seeds, and targeted ablations isolating the roles of attention, residual paths, and gated fusion. The results indicate that SAMRBNET favors globally consistent reconstructions with reduced ringing/checkerboard artifacts, while remaining efficient and reproducible.

Predicting the energy consumption of buildings plays a critical role in supporting utility providers, users, and facility managers in minimizing energy waste and optimizing operational efficiency. However, this prediction becomes difficult because of the limited availability of supervised labeled data to train Artificial Intelligence (AI) models. This data availability becomes either expensive or difficult due to privacy protection. To overcome the scarcity of balanced labeled data, semi-supervised learning utilizes extensive unlabeled data. Motivated by this, we propose semi-supervised learning to train AI model. For the AI model, we employ the Belief Rule-Based Expert System (BRBES) because of its domain knowledge-based prediction and uncertainty handling mechanism. For improved accuracy of the BRBES, we utilize initial labeled data to optimize BRBES’ parameters and structure through evolutionary learning until its accuracy reaches the confidence threshold. As semi-supervised learning, we employ a self-training model to assign pseudo-labels, predicted by the BRBES, to unlabeled data generated through weak and strong augmentation. We reoptimize the BRBES with labeled and pseudo-labeled data, resulting in a semi-supervised BRBES. Finally, this semi-supervised BRBES explains its prediction to the end-user in nontechnical human language, resulting in a trust relationship. To validate our proposed semi-supervised explainable BRBES framework, a case study based on Skellefteå, Sweden, is used to predict and explain energy consumption of buildings. Experimental results show 20 ± 0.71% higher accuracy of the semi-supervised BRBES than state-of-the-art semi-supervised machine learning models. Moreover, the semi-supervised BRBES framework turns out to be 29 ± 0.67% more explainable than these semi-supervised machine learning models.

The rapid expansion of Bangladesh’s digital economy has unlocked new opportunities for online businesses, but it has also heightened the challenges of delivering high-quality e-services in an increasingly competitive environment. While previous studies have largely focused on customer satisfaction and consumer experiences, the perspective of business owners—those managing daily operations and strategic decisions—has been notably underexplored. Addressing this gap, the present study aims to identify and prioritize the critical success factors (CSFs) influencing the performance of online businesses in Bangladesh from the owners’ viewpoint.A structured survey has been conducted among 57 online business owners, collecting data across ten essential dimensions, including website infrastructure, information sharing, delivery practices, responsiveness, customization, customer service, reliability, payment security, return policies, and service quality challenges. To ensure a data-driven analysis, the Random Forest algorithm is utilized for feature importance extraction, followed by pairwise comparisons through fuzzy triangular matrices. Subsequently, FAHP and TOPSIS techniques are applied to systematically rank the factors and alternatives. Model validation using the XGBoost classifier results in 93% accuracy and a 93% F1-score, demonstrating the robustness of the proposed approach. The results reveal that service quality concerns, payment security, and effective customer support are the most influential determinants of business success. Beyond offering a methodological advancement that integrates machine learning with MCDM, this study delivers actionable insights for business owners, digital strategists, and policymakers aiming to enhance the competitiveness and resilience of Bangladesh’s e-commerce landscape.

Speech Emotion Recognition (SER) is a complex endeavor in human-computer interaction (HCI) that necessitates the use of artificial intelligence and deep learning to accurately classify emotional states, which are determined by analyzing speech audio signals. Bangla is classified as a low-resource language for SER due to the scarcity of labeled datasets, despite its status as the seventh most frequently spoken language at the global level. By utilizing the SUBESCO and BanglaSER corpora, which are both audio-only Bangla emotive speech datasets, this study aims to enhance emotion recognition in Bengali speech. Noise was eliminated through the application of Envelope Masking during preprocessing, and Mel-Frequency Cepstral Coefficients (MFCCs) were extracted to capture critical spectral features. Machine learning models such as K-Nearest Neighbor (KNN), Random Forest, and Multi-Layer Perceptron (MLP) are implemented by the system, in addition to ensemble techniques like Voting and Stacking Classifiers, to optimize its performance. Further, in order to process temporal and sequential speech patterns efficiently, deep learning architectures such as Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) were implemented. The models that were proposed exhibited a high level of perceptual efficiency, obtaining an accuracy of 95.92% on SUBESCO and 90.61% on BanglaSER. These findings substantiate the efficacy of the preprocessing techniques and applied learning models, thereby enhancing BanglaSER and broadening the scope of research opportunities for low-resource languages.

Electricity generation in Bangladesh is predominantly from fossil fuel power plants. Bangladesh’s electricity generation from renewable sources is well below the global average, with less than 1% coming from hydropower and less than 1% from solar and wind. Estimating and reducing carbon emissions from gas, coal, and oil generation plants in Bangladesh is a burning issue. The carbon emission factor is used to compute the carbon dioxide emissions from fossil fuel plants based on their generation capacity, and fuel type. The Intergovernmental Panel on Climate Change (IPCC) framework is used to calculate the carbon emission factor of each fuel type. The paper presents Matpower Interior Point Solver (MIPS), and Dynamic Non-Linear Particle Swarm Optimization (DNPSO) algorithms to solve combined economic emission dispatch (CEED) to optimize generation dispatch and minimize emissions. A practical 14-bus test system in Southern Chattogram, consisting of 132 kV, 230 kV, 400 kV, and 765 kV lines, is constructed. The developed algorithms based on optimal power flow have applied on 14-bus test system and measured optimal dispatch and emissions of the generators and grid. The environmental emission factor is used to combine the economic and emission dispatch. The existing and future coal power plants at Matarbari will release from 1934 ton-CO2 to 4889 ton-CO2 in year 2025 to 2041 with current plant efficiency. The results indicate that enhancing plant efficiency to 47%, the carbon dioxide emissions from coal power plants may be decreased to 15%. Moreover, when CEED is implemented with MIPS and DNPSO, the daily carbon emissions from the 14-bus test system are reduced to 14%, while those from DNPSO decreased to 11% from their original levels. CEED prioritizes more cleaner generators in the system than higher emissions low-cost generators. Furthermore, transmission line losses and grid emissions are lower in economic-emission dispatch compared to economic dispatch.