y
a
CONFIRM: Comprehensive, Orchestrated, Non-Negotiable Framework for Information Security and Risk Management
This research contributes “CONFIRM (Comprehensive Orchestrated, Non-Negotiable Governance Framework for Information Security and Risks Management)” as an integrated governance framework designed to eliminate information-security governance blind spaces through continuous end to end visibility and traceability. The framework establishes a traceable relationship from “Organizational Vision → Mission → Objectives → Goals → Activities → Evidence/Status”, while integrating “people, processes, technologies, enterprise assets, risks, controls, compliance obligations, security operations, and resilience” within a unified governance architecture. Its principal contribution is the transformation of fragmented operational and security information into continuously traceable governance intelligence through “Enterprise Relationship Intelligence (ERI)” and a “Continuous Governance Intelligence (CGI)” cycle. Unlike conventional point-in-time or siloed governance approaches, CONFIRM enables responsible and accountable parties, management, the Board, and relevant authorities to determine the current status, ownership, dependencies, risks, control effectiveness, evidence, and business impact of governance activities and to trace these upward to organizational objectives. The framework further introduces “automated multi layer processing, quantitative governance/risk/resilience measurement, and AI embedded decision support”, enabling early identification of deviations, evidence based decisions, timely corrective actions, and measurable assessment of whether information security activities along with other operational and business activities are effectively supporting organizational objectives and the achievement of the organizational Vision.
Privacy-Preserving Single-Lead Smartwatch ECG Monitoring for Multi-Class Arrhythmia Detection
This work presents a fully local, privacy-preserving inference framework, ECGNet, for multi-class arrhythmia detection from simulated 1-lead smartwatch ECG, extending beyond binary Normal/AFib detection to four clinically relevant rhythm classes. A confidence-calibration mechanism withholds low-confidence predictions rather than forcing unreliable diagnoses, and a fair 12-lead vs. single-lead benchmark under an identical model quantifies the diagnostic cost of wearable-grade acquisition, achieving 90.64% accuracy with a 9.36% inconclusive rate at a calibrated 0.80 confidence threshold. A hardware-agnostic sampling-rate simulator further evaluates robustness across device-specific acquisition rates without requiring physical wearable hardware.
Multi-Backbone Ensemble Learning for Robust Breast Cancer Classification on the Mini-DDSM Mammography Dataset
This study presents a rigorous patient-grouped evaluation framework for breast cancer classification on the Mini-DDSM dataset, while identifying and addressing cross-folder patient-identifier collisions that can distort patient grouping and fold composition. It systematically evaluates EfficientNet-B4 and ConvNeXtV2-Tiny using test-time augmentation, threshold analysis, and Grad-CAM explainability, providing a reproducible reference framework for robust mammography classification.
OTSM-Configured Strategic Communications for Integrated 6G Autonomous Metro Rail Networks
The paper introduces a strategic communication backbone tailored specifically for integrated 6G autonomous metro rail networks. By modeling the complex integration of wireless sensors directly into the metro rail’s core communication framework, the research provides a targeted architectural solution designed to minimize latency in mission-critical, real-time metro communications.
Fabrication and Performance Analysis of a Zinc-Bromine Single Flow Battery
A low-cost single-flow zinc-bromine cell is fabricated in which the positive-side reservoir and pump are eliminated, using non-woven glass fiber as the separator – a material established in lithium-based and lead-acid systems together with a graphite positive electrode backed by porous carbon felt to increase active surface area and retain evolved bromine.
Off-Grid Electrification of Urir Char Using HOMER-Optimized Hybrid Renewable Microgrids
Addressing the critical challenge of off-grid island electrification in Bangladesh, this study presents the first comprehensive HOMER Pro techno-economic and environmental evaluation of a hybrid renewable microgrid tailored for Urir Char. The primary contribution lies in the novel integration of solar PV, hydrokinetic turbines harnessing strong estuarine water currents, biogas generators utilizing local agricultural and animal waste, and Lithium Iron Phosphate battery storage, a system architecture previously unexamined in HOMER Pro literature for Bangladeshi off-grid coastal regions. The study demonstrates that an optimal 100% renewable Load Following configuration (PV/Biogas/Hydrokinetic/Battery) delivers reliable energy access to 16,078 residents at an exceptionally low Levelized Cost of Energy (LCOE) of $0.0629/kWh and a Net Present Cost (NPC) of $3,134,828. Furthermore, the findings prove that a locally fueled biogas generator operating for just 353 hours per year effectively eliminates the high economic, logistical, and environmental risks associated with diesel transport while maintaining near-zero carbon emissions (507 kg/year), highlighting that dispatchable bio-energy, rather than over-sizing solar or battery capacity, is essential to resolving renewable intermittency.
Effects of Categorical Encoding, Normalization, and Feature Reduction on Ensemble Learning for Mushroom Edibility Classification
Mushroom edibility classification is an important problem because poisonous mushrooms may cause serious health risks. Machine learning methods can be useful for classifying mushrooms using their different characteristics. However, preprocessing methods and the number of features can affect classification performance. In this study, the effects of categorical encoding, normalization, and feature reduction are investigated using the Secondary Mushroom dataset. Two categorical encoding methods, label encoding and frequency encoding, are considered. Min-max normalization is also applied to examine its effect on the results. Four ensemble machine learning models, namely Random Forest, XGBoost, LightGBM, and AdaBoost, are evaluated. The features are ranked according to their importance, and the models are tested using the top 2 to 20 features. The experiments use an 80:20 hold-out split with a fixed random state. The results show that frequency encoding improves the performance of XGBoost and LightGBM for different feature subsets. Normalization produces almost no change in the accuracy and feature ranking of the investigated tree-based models. Random Forest and LightGBM achieve at least 99.9% accuracy using only eight features. The results indicate that categorical encoding and feature reduction can affect the performance and feature requirement of ensemble models for mushroom classification.
LiteDR-XAI: A Lightweight and Explainable Framework for Referable Diabetic Retinopathy Screening
Automated diabetic retinopathy screening must balance predictive performance with computational cost, threshold control, reproducibility, and interpretability. This paper presents LiteDR-XAI as a reproducible framework for binary referable diabetic retinopathy screening on APTOS 2019, not as a new con volutional architecture or a universal state-of-the-art model. Five pretrained backbones are evaluated through 35 controlled runs covering input resolution, contrast-limited adaptive histogram equalization, pretraining, backbone freezing, and loss design. Checkpoints are selected by validation AUROC, and the operating threshold is estimated only from validation predictions using Youden’s index before application to the held-out test set. From these experiments, a compact MobileNetV3-Large configuration at 320 pixels is selected as the LiteDR-XAI operating model because it provides a balanced accuracy-efficiency-explainability trade-off: 94.73% accuracy, 0.9791 AUROC, 0.9646 AUPRC, 95.96% sensitivity, 93.88% specificity, 91.45% precision, and 93.65% F1-score with 3.22 million parameters. A strict literature comparison includes only near-direct APTOS binary/referable DR studies and shows that LiteDR-XAI has the highest accuracy among the included comparisons, while other methods report higher AUC or sensitivity under different protocols. Grad CAM supports qualitative inspection of correct and erroneous predictions. External, multi-seed, device-level, and prospective validation remain necessary before deployment.
A Rule-based Automatic Algorithm for Detecting Total Sleep Time using Pulse Oximetry Signal
This study proposed a lightweight and interpretable rule-based framework to estimate TST from overnight finger-tip pulse oximetry signals only. After pre-processing, different features were extracted from each 30 second epoch and the rule-based framework classified each epoch as either sleep or wake. After applying TST correction, the framework finally estimated the TST as the output. Performance evaluation demonstrated significant correlation between estimated and annotated TST values (Pearson’s correlation coefficient, r = 0.804 and Lin’s concordance correlation coefficient, CCC = 0.760). Bland-Altman plot demonstrated small mean bias (0.24 hours) with a narrow limits of agreement (-0.95 to 1.44 hours). The proposed framework provides reliable and computationally efficient TST estimation, thus a promising approach for portable and home-based sleep monitoring applications.
Technological Innovation in Precision Agriculture: An EfficientNet-B0 Transfer Learning Framework for Grape Leaf Disease Diagnosis
This paper advances mobile precision agriculture by engineering an optimized EfficientNet-B0 transfer learning framework that resolves the critical trade-off between high diagnostic precision and edge-compute efficiency in vineyard biosecurity. Unlike legacy architectures like VGG16 that suffer from high parameter redundancy (138M+ parameters) and slow inference latency, our compound-scaled model achieves a peak 94.80% diagnostic accuracy and 94.80% F1-score across field-collected grape foliage classes while maintaining a remarkably lightweight profile of 5.3 million parameters, a 21 MB memory footprint, and an execution cost of just 0.39 GFLOPs. Furthermore, by evaluating performance directly on outdoor, variable-field imagery rather than controlled laboratory benchmarks, this work demonstrates a practical, parameter-efficient pipeline tailored for real-time, offline edge-AI deployment on resource-constrained mobile hardware and autonomous agricultural drones.
TRACE-BN: Transferring Bangla-English Tutoring Behavior to a Sub-1B Offline Language Model
This paper introduces TRACE-BN, a curriculum-guided bilingual tutoring dataset of structured traces for Bangla-to-English language learners at the CEFR A1–A2 level. Each trace encodes a complete seven-field pedagogical sequence combining word-level glosses, literal and natural translations, contrastive Bangla grammar explanations, common learner mistakes, and targeted practice with answers. We show that a compact sub-1B offline model (Qwen3-0.6B) can acquire this multi-component tutoring behavior via 4-bit LoRA fine-tuning, improving schema validity from 85.4% to 95.8% while raising translation and pedagogical scores. We validate the approach through an expert human audit of the supervision signal and a reference-aware dual-judge evaluation showing consistent improvements across all evaluated tutoring dimensions.
SafeNet-Guard: A Secure Reporting Platform for Technology-Facilitated Harassment with Evidence Integrity and Duress Signaling
Four contributions follow from our proposed design: (i) a role-stratified secure reporting architecture, (ii) tamper-evident versioning of reports and evidence, (iii) a dual-credential coercion-response protocol, and (iv) a controlled evaluation that stress-tests these components against tampering and coercion scenarios.
First-Principles Study of Lead-Free Cs2NaGaBr6-xFx (x = 0.00, 0.22, 0.33) Perovskites for Next-Generation Functional Materials
In this work, a systematic density functional theory study on lead-free double perovskites Cs₂NaGaBr₆₋ₓFₓ (x = 0.00, 0.22, 0.33) is reported, which reveals the changes brought by F substitution on the structure, electronic, optical, mechanical and elastic properties of these compounds. This paper finds out the variations in electronic structure and band gap parameters along with the enhancement in optical response and absorption of these compounds, which suggests that halide engineering can play an important role in tuning the functionalities of these materials. The discussion on the elastic and mechanical properties of the materials helps in understanding the structural stability, mechanical strength for application of next generation materials.
Comparative Techno-Economic Analysis of Si, GaAs, and InGaN Thin-Film Solar Cells under Bangladesh’s Climatic Conditions
The performance of photovoltaic (PV) systems is influenced by operating temperature, especially in Bangladesh’s tropical climate. This study presents a comparative numerical investigation of Si, GaAs, and InGaN p-i-n thin-film solar cells using site-specific meteorological data. Temperature-dependent photovoltaic characteristics were investigated using SCAPS-1D simulation software. Quantum efficiency (QE) and spectral response (SR) were evaluated to characterize optical absorption performance across the AM1.5G solar spectrum. GaAs achieves the highest power conversion efficiency (PCE) of 23.61% due to its superior thermal stability. InGaN demonstrates potential for high-temperature environments, yielding the highest open-circuit voltage and a stable fill factor due to its wide tunable bandgap. In contrast, silicon exhibited the highest thermal sensitivity and the lowest PCE (10.51%) due to the lower light absorption of the thin absorber layer. Furthermore, economic feasibility assessment using RETScreen for a proposed 10 MW utility-scale power plant revealed that GaAs-based system requires the smallest solar collector area (41,391 m²) and generates 16,530.91 MWh annually at the lowest cost of electricity ($0.0777/kWh). In contrast, the silicon-based system requires the largest collector area and incurs the highest energy production cost.
CyberMirror: Recovering Interpretable Attacker Reward Functions from Deep-RL Pentesting Agents via Inverse Reinforcement Learning
1. A demonstration-free pipeline on an emulation-friendly
range that transforms a PPO/MLP attacker into an
attacker-value detector without any human data, using
a sample-based Maximum-Entropy IRL reward recovery
algorithm with a separable objective and real partition
function, made to converge by (i) Markovian state-
occupancy features, (ii) horizon-matching the background
to the median demonstration length, and (iii) Ridge
regularization (λ = 0.3).
2. Empirical evidence that the recovered score differentiates
adversarial from benign states (AUC ≈ 1.0, 100% detect-
before-goal), and a detection-gated, shared-environment
duel in which this reward incentivizes an action-taking
defender, reported with its availability cost so that con-
tainment is not mixed up with free defense.
3. A transferable methodological finding on the trade-off be-
tween attack optimality and reward identifiability in IRL-
from-RL, explicitly stating the limitations and providing
a future roadmap.
Exploring SVM and CNN Behavior Across Kepler Optical Light Curves and TART Radio Visibilities
The research study is organized around the following objectives and contributions:
• Formulate a shared binary quality-control experiment that distinguishes original observations from deterministically corrupted copies using four analogous, modality-specific corruption categories at three severity levels.
• Construct provenance-tracked, source-disjoint Kepler and TART workflows with an equal-condition stress grid applied to every held-out source.
• Explore SVM and CNN behavior through a matched cross-modality design using balanced test metrics, source-level bootstrap intervals, and a 60-session TART sensitivity analysis, without claiming novelty for the individual algorithms or corruption concepts.
Operational Assessment of a Small-Scale Solar PV System : A Case Study from Bangladesh
The goals are to: (i) validate
the consistency of the solar datasets; (ii) quantify the daily/period
yield, FLH, specific yield and ca-capacity factor; (iii) test the
dataset sensitivity with six categories of the logged weather
data; and (iv) position the results in the context of the regulatory
framework of SREDA and other similar, peer-reviewed regional
studies
Social Media Sentiment Analysis for Local Businesses
This study develops a unified Yelp-based sentiment analysis framework that systematically compares lexicon-based, traditional machine-learning, and transformer-based models under a common evaluation setting. The best-performing model, DistilBERT, is then applied to a separate 300,000-review sample to support category-, aspect-, and business-level analysis. The main contribution is the integration of sentiment classification with keyword-based aspect detection and category-relative benchmarking to identify interpretable business strengths, weaknesses, and actionable customer-feedback patterns.
Explainable and Fair Student Dropout Prediction from Multimodal Temporal Learning Analytics
We propose a multimodal temporal framework that integrates demographic, academic, and week-by-week VLE behavioral data for early student dropout prediction.
We develop a dual-input BiLSTM architecture that jointly models static student characteristics and temporal learning engagement patterns to improve predictive performance.
We incorporate SHAP and LIME to provide both global and student-level explanations, enabling educators to understand the key factors underlying dropout predictions.
We conduct a systematic fairness assessment across gender, age, disability status, and socio-economic deprivation to identify potential disparities in model performance and prediction errors.
We translate predictive and explainability outputs into risk levels and personalized intervention recommendations, supporting actionable early-warning decision-making in virtual learning environments.
Real-Time Blind Spot Vehicle Detection and Risk Classification System Using Deep Neural Network
This study is about a way to detect things around autonomous vehicles in real time. It uses a two step process. First it finds objects using YOLOv8. Then it figures out what those objects are using other methods like EfficientNetV2, MobileNetV2, VGG16 and ResNet-50. The study used a lot of pictures of roads around 9,850 of them to make this system work. EfficientNetV2 was the best at getting it right it was correct 91 percent of the time and it was very good at finding things that could be dangerous. This system is as good as or even better than systems that use radar. It only needs one low cost camera. The study also looked at how these methods work for detecting blind spots, which is a big problem on highways in Bangladesh where there are a lot of different types of vehicles on the road and accidents happen because people cannot see what is, around them.
An Empirical Comparison of Monolithic and Microservices Architectures for an E-Commerce Application
This paper presents an empirical performance comparison of monolithic and microservices architectures for an e-commerce application. Using k6 load testing at 50 and 100 virtual users, we demonstrate that microservices achieve 5.4% higher throughput, 25% lower average latency, and 39% lower p95 latency at 100 VUs, while exhibiting different failure modes. The monolith shows consistent order-creation failures (~0.7% error rate), whereas microservices failures are transient and service-specific. This study provides quantitative evidence that architectural decomposition can improve scalability and tail latency under stress, with implications for software architecture decision-making.
A Comparative Analysis of Multimodal Frameworks for Multilingual Human vs. AI Caption Detection
This paper presents a comprehensive comparative analysis of multimodal frameworks for detecting human vs. AI-generated captions in a multilingual context. It identifies the most robust framework architectures and provides critical insights into cross-lingual detection accuracy, helping improve the reliability of AI-generated content auditing.
MedlmBD: An AI-Driven Framework for Bengali Healthcare Query Management
• Designing a medical AI assistant that addresses key gaps in existing research
• Developing and fine-tuning the SmolLM3-3B model using a subset of the ChatDoctor Dataset
• Performing a comprehensive evaluation and comparison with a pre-trained medical domain model for evaluating the quantitative and qualitative metrics of our architecture
• Overcoming language barriers in healthcare support through the use of a Bengali-English translation pipeline.
• Despite using only a fraction of the parameters and being fine-tuned only on a subset of the ChatDoctor dataset, fine-tuned SmolLM3-3B is competitive with the larger MedAlpaca-7b model.
CropGuard AI: A Crop Disease Diagnosis System Combining EfficientNet-B0 and Vision-Language Models for Bangladeshi Agriculture
The research contributes an integrated crop disease diagnosis framework that combines EfficientNet-B0 image classification with vision-language diagnosis, Bangladesh-focused disease knowledge, and RAG-based agricultural advisory to provide more comprehensive and accessible crop disease support.
A Leakage-Aware Ordinal Machine Learning and Explainable AI Framework for Hidden Replication, Label Ambiguity, and Performance Inflation in Maternal Health Risk Prediction
This paper proposes four significant contributions:
1) We quantify the scale of duplication and label ambiguity and its practical consequence for standard train/test splitting.
2) We directly measure the resulting performance inflation by comparing conventional cross-validation against exact-row-grouped and physiological-profile-grouped cross-validation for two widely used models.
3) We benchmark a broad family of ordinal-aware and conventional classifiers under leakage-aware, repeated nested cross-validation, including a two-threshold ordinal decomposition and a CORAL-style ordinal neural network, with full calibration and conformal-prediction analysis.
4) We evaluate the stability of SHAP-based explanations across random seeds, treating explanation reliability itself as an empirical outcome rather than an illustrative afterthought.
An Explainable and Robust Machine Learning Framework for IoT Intrusion Detection through Zero-Day and Cross-Dataset Evaluation
This research proposes a comprehensive and repeatable machine learning system for IoT IDS, covering leakage-aware pre-processing, comparison of five ML models, class imbalance, explainability using SHAP and a lightweight 15 feature IDS. The framework also assesses the robustness to zero day attacks under a Leave-One-Attack-Out approach and compares the supervised detection approach with an Isolation Forest baseline. Furthermore, bidirectional cross-dataset evaluation between TON_IoT and UNSW_NB15 is conducted to evaluate the model in different network environments. The results show that explainability, resistance to unseen attacks, and cross-dataset validation are necessary to measure the usability of intrusion detection systems in IoT.
A Hybrid Multi-Backbone Deep Learning Framework with Learnable Attention Fusion for Explainable Potato Leaf Disease Detection
Potato leaf diseases, particularly early blight and late blight, significantly reduce crop yield and quality, making accurate and timely diagnosis essential. This paper introduces a hybrid multi-backbone deep learning framework, which combines EfficientNet-V2-S, Swin Transformer and ViT-Base, and introduces a Learnable Attention Fusion (LAF) module to capture complementary local and global feature representations. A progressive layer-freezing approach is used to enhance transfer learning and model generalization. The proposed framework is tested in terms of 5-fold cross validation and cross dataset testing. The experimental results are obtained with good average classification accuracy of 99.47% for 5-fold cross validation and 97.34% for cross-dataset evaluation, exhibiting good robustness and generalization capability for different data distributions. Moreover, Grad-CAM is used to give visual interpretation of model predictions by identifying the relevant areas of the image associated with the disease. The proposed model is used as a web application for practical agricultural applications, allowing identification of potato leaf diseases in real-time based on uploaded leaf images. From the experimental results, it is proved that the proposed framework gives an accurate, robust and interpretable solution for the potato leaf disease automatic classification.
Multi-Class Brain Tumor MRI Classification Using EfficientNet Based Hybrid Attention Network
Brain tumor diagnosis from magnetic resonance imaging (MRI) plays an important role in supporting early clinical assessment. However, distinguishing different tumor types remains challenging because of variations in tumor size, shape, and appearance across patients. This paper presents an enhanced EfficientNetV2B0-based model that combines the Convolutional Block Attention Module (CBAM) with Multi-Head Self-Attention (MHSA) to improve feature representation for multi-class brain tumor classification. The model is evaluated on a publicly available dataset containing 7,200 brain MRI images from four categories: glioma, meningioma, pituitary tumor, and no tumor. Its performance is compared with EfficientNetV2B0, EfficientNetV2B0 with CBAM, EfficientNetV2B0 with MHSA, ResNet50, DenseNet121, and MobileNetV3Small under the same experi- mental setting. The proposed model achieved a test accuracy of 96.30%, with a weighted F1-score of 96.29%, a mean Dice score of 96.36%, and a mean IoU of 93.02%, outperforming all baseline models. It also obtained macro and micro AUC values of 99.65% and 99.63%, respectively, demonstrating consistent clas- sification performance across all classes. These findings indicate that integrating channel attention with self-attention helps the network learn more discriminative features while preserving the computational efficiency of the EfficientNetV2B0 backbone.
Renewable Energy Consumption and Sustainable Wealth in Developing Countries: The Moderating Role of Government Effectiveness
These outcomes highlight the importance of strengthening institutional quality alongside renewable energy expansion to promote sustainable wealth accumulation in developing countries.
Graph-Based Clustering and Explainable Graph Neural Networks for Crime Hotspot Prediction in Bangladesh (2020–2025)
Bangladesh has experienced noticeable variation in crime across its metropolitan and range-level policing units during 2020–2025. Some regions, such as Dhaka and Chattogram, regularly report high volumes of violent and property crimes, while others show more sporadic spikes. Conventional analysis tools, for example, trend plots or static heatmaps, often describe these patterns in isolation and struggle to capture how crime in one area resembles or relates to crime in another. This paper presents a graph-based framework that combines community detection, Graph Neural Networks (GNNs), traditional machine learning, and explainable AI (XAI) to anticipate high-risk regions and to understand how policing units are connected in terms of crime behaviour. Monthly crime data from 17 units are used to construct a similarity graph where nodes represent units and edges represent similarity in crime statistics. Louvain and spectral clustering reveal three stable crime-behaviour communities, and the same engineered features feed Random Forest, Graph Convolutional Network (GCN), and Graph Attention Network (GAT) models for next-month high-risk prediction. Random Forest achieves the highest overall accuracy, while the GAT model attains the highest recall on high-risk months, making it useful where missing an emerging hotspot is costlier than a false alarm. SHAP and GNN Explainer are applied to interpret the Random Forest and GCN respectively, showing that Random Forest concentrates its decisions on a small set of temporal features while the GCN distributes importance more evenly across features and relies on a specific subset of high-weight connections between policing units (importance-ranking correlation between the two methods = 0.5125). The results indicate that graph-based analysis, combined with explainability tools, can make crime forecasts more interpretable and useful for planning, even when a classic tabular model remains the most accurate overall predictor.
A Machine Learning Approach to Screening for Child Undernutrition in Rural Bangladesh
This paper builds separate stunting, underweight, and wasting classifiers for rural under-five children from BDHS 2022, correcting class imbalance strictly inside a leakage-free training pipeline (SMOTE and class weighting applied only after the train–test split) and using recall as the primary metric. A class-weighted logistic regression gives the best recall on all three indicators and detects roughly twice as many wasted and underweight children as the closest published model on the same survey. A high-accuracy random forest that identifies none of the wasted children is shown to demonstrate why recall must be prioritised over accuracy on this imbalanced data.
Reference-Conditioned Healthy-Like EEG Reconstruction Using Movement-Aware CycleGAN for Motor Imagery BCI
Proposes a movement-aware, reference-conditioned CycleGAN to reconstruct healthy-like EEG from stroke motor-imagery EEG using movement-matched healthy references and anti-copy and movement-consistency constraints, with subject-level holdout evaluation to reduce data leakage and separately assess healthy-domain similarity and movement preservation.
Domain-Aware Hybrid Quantum Learning via Correlation-Guided Circuit Design for Crime Pattern Analytics
– Developed a comprehensive quantum-classical comparison for crime analytics with statistical validation through cross-validation across quantum, classical, and hybrid paradigms.
– Proposed a novel quantum circuit architecture that utilizes crime feature correlations through targeted entanglement based on Spearman correlation analysis. — Implemented hybrid integration strategies, including Q→C (quantum feature extraction followed by classical classification) and C→Q (classical dimensionality reduction followed by quantum modeling).
Addressing Bandwidth and Isolation Challenges in THz MIMO: A Dual-Band Bow-Tie Antenna for Biomedical, Pharmaceutical, and Security Applications
Terahertz (THz) MIMO antenna systems offer strong potential for next-generation sensing, imaging, and wireless communication, but practical deployment is still limited by narrow impedance bandwidth and poor inter-element isolation in compact designs. This work presents a dual-element MIMO antenna on a quartz substrate (170 × 68 µm²) that addresses both challenges using a modified bow-tie radiator with embedded rectangular slots. The slot-loading creates independent resonant paths within a single element, enabling dual-band operation at 5.12 THz and 6.78 THz without increasing footprint. Full-wave simulations (CST Microwave Studio) show a combined −10 dB bandwidth of 757 GHz, with return losses of −40.64 dB and −35.5 dB and VSWR values of 1.01 and 1.03, indicating near-ideal impedance matching. The antenna achieves peak directivities of 6.22 dBi and 7.44 dBi with stable linear polarization. Inter-element isolation remains better than −30 dB across the band without any decoupling structures. MIMO performance metrics further confirm strong diversity behavior, with an ECC of 1.25 × 10⁻⁷, CCL of 6.84 × 10⁻¹⁵ bps/Hz, and diversity gain of 9.99 dB, close to the theoretical limit. The proposed design serves as a compact, fabrication-compatible platform for biomedical imaging, pharmaceutical spectroscopy, security screening, and 6G MIMO communication within a single THz antenna architecture.
Design and Analysis of LP01 Mode Hollow-Core Anti-Resonant Fiber with Hybrid Elliptical Geometry
This research endeavor represents a novel air filled anti-resonant hollow-core fiber (HC-ARF) designed to achieve ultra-low transmission loss as well as efficient LP₀₁ (single-mode) guidance. The fiber geometry is fine-tuned through successive parameter sweeps. This reduces the interaction between the guided optical field and the surrounding silica significantly, which lowers the propagation loss and improves the transmission performance. The optical characteristics of the proposed fiber are analyzed using FEM (Finite Element Method). The simulation results demonstrate an ultra-low confinement loss of 4.16 × 10⁻⁶ dB/m over the wavelength range of 1.4–1.7 μm. In addition, the fiber attains a Higher Order Mode Extinction Ratio (HOMER) approaching 100 which indicates the effective suppression of higher order modes and stable single mode operation. The bending performance is also excellent, with a value of only 7.35 × 10⁻⁵ dB/m at the standard telecommunication wavelength of 1.55 μm for bending radii larger than 10 cm. These results demonstrate that the proposed HC-ARF achieves low loss and high mode purity while retaining good bending resistance, all within a simple structural design. Its low loss and modal stability suit it to long haul optical communication and high speed data transmission, whereas the hollow core itself opens further use in optical sensing and high-power laser delivery around 1.55 μm.
High-Efficiency Lead-Free Cs2PdBr6 Perovskite Solar Cells: A SCAPS-1D Numerical Study of Transport Layer Optimization
In this study, a lead-free Cs2PdBr6-based per-ovskite solar cell (PSC) has been numerically simulated systematically with help of SCAPS-1D software with an FTO/ETL/Cs2PdBr6/HTL/Au architecture. Three electron trans-port materials (WO3, TiO2, and SnS2) and three hole transport materials (MoO3, CuI, and GO) were comparatively evaluated across varying layer thicknesses. The optimum thicknesses of the FTO and absorber layers are determines to be 0.01 μm and 5 μm, respectively. Among the investigated transport materials, WO3 and MoO3 used as the electron transport layer (ETL) and the hole transport layer (HTL), respectively, due to their favorable energy band alignment and efficient charge carrier extraction capability. The combined effect of absorber defect density (1010 to 1020 cm−3) and operating temperature (280 K to 360 K) was further analyzed, confirming that crystal quality is the dominant performance-limiting factor. The fully optimized
FTO/WO3/Cs2PdBr6/MoO3/Au device achieves a PCE of 30.62%,Voc = 1.30 V, Jsc = 27.29 mA/cm2, and FF = 86.31% at 300 K, sur-passing all previously reported Cs2PdBr6-based configurations.
MediLink: Design and Validation of a Portable IoT System for Physiological Monitoring
The major contribution of this research is the development of MediLink, a portable ESP32-S3-based IoT platform for simultaneous monitoring of HR, SpO₂, temperature, and single-lead ECG. The system integrates local visualization, GPS tracking, and cloud-based remote monitoring using a FreeRTOS-based architecture. The study also provides a preliminary evaluation against reference devices and discusses the limitations of the validation process.
CALIBER: Pre-Generation Gated Image-Exemplar-Backed Retrieval-Augmented Generation for Trustworthy Dermoscopic Decision Support
The significant research contribution is the CALIBER framework, a calibrated visual-exemplar retrieval-augmented pipeline for dermoscopic decision support that combines DermCLIP-based retrieval, evidence sufficiency checking, conformal prediction, diagnostic confidence thresholds, and abstention. It improves not only diagnostic accuracy but also calibration, faithfulness, and evidence-grounded clinical reasoning, achieving 89.2% accuracy and demonstrating more trustworthy VLM-assisted dermatology decisions.
Performance Evaluation of a DeMod-CDSC-Based P-Class PMU Under Noisy Conditions
The key contribution is the first systematic noise-robustness assessment of the frequency-adaptive DeMod-CDSC P-class PMU framework, demonstrating superior TVE and FE performance over SOGI-CDSC across 30–80 dB SNR and establishing its suitability for noisy real-world smart-grid applications.
TCA-Net: A Tri-Stage Cross-Attention Network with Explainability-Consistency and Boundary-Aware Learning for Generalizable Medical Image Segmentation
Existing medical image segmentation methods often address cross-site generalization, boundary accuracy, and explainability independently, leaving a gap in their unified integration. To address this, we propose TCA-Net, a novel CNN–Transformer framework that combines four-level Tri-Stage Cross-Attention fusion with training-time XAI-Consistency learning and boundary-aware supervision. Unlike conventional feature-fusion methods, TCA-Net explicitly captures complementary channel, spatial, and cross-attention information while promoting anatomically meaningful and precise segmentation. The framework is validated using leakage-free cross-site evaluation and controlled ablation studies, achieving 96.28% Dice in-domain and up to 98.29% Dice across sites.
IoT-Powered Hydroponic Tomato Monitoring System
This thesis demonstrates innovative use of IoT technology in hydroponic tomato cultivation, enabling real-time monitoring and control, which significantly advances precision agriculture.
Performance Evaluation of Contrast Enhancement Techniques for YOLOv8-Based Fall Detection under Low-Light Conditions
The detection of falls in low-light indoor environments is a challenge for vision-based systems because of the low visibility of features and degradation of contrast. Deep learning techniques with the YOLO algorithm provide an alternative to contact-based measurements but struggle under poor lighting conditions, and existing solutions are based on expensive IR or thermal cameras. In this study, a structured evaluation framework is proposed to systematically study the effect of contrast enhancement on fall detection with conventional RGB cameras in a controlled low-light indoor environment. Three enhancement techniques, namely Histogram Equalization (HE), Adaptive Equalization (AE), and Contrast Stretching (CS), are tested under various lighting conditions. The results show the consistent and systematic relationship between contrast restoration and the detection performance, with global enhancement showing more stable and precise results. Higher accuracy is obtained in darker scenes (75%) and dimly lit scenes (62.5%), with a precision of 1.000 in darker scenes with a high level of accuracy and thus reducing false alarms. The proposed framework can generate reproducible evidence for improving low-light detection performance and does not need any special imaging hardware or increase the complexity of models, which can promote the development of stable and scalable indoor fall detection systems for real applications.
Context-Aware E-Commerce Recommendation via Knowledge Graph- Enriched Graph Neural Networks
We chose a knowledge-graph-plus-GNN approach
specifically because it addresses this gap without abandoning
collaborative signals altogether. Rather than replacing user-
item interaction data with content-based rules, the proposed
pipeline keeps the interaction graph as the backbone and
augments it with structured attribute edges (Item–Brand,
Item–Category), allowing the GNN to propagate preference
signal along attribute pathways that a plain bipartite model
cannot see. Neo4j was chosen as the underlying store because
it allows this structure to be expressed, queried, and inspected
directly as a graph via Cypher and the Graph Data Science
library rather than reconstructed implicitly inside a black-box
model, which also makes the resulting recommendations
easier to explain to a non-technical stakeholder (Section V-E
gives a concrete example of this). Critically, we did not
simply assume that adding this context would help; the
ablation design in Section V-E was built specifically to test
that assumption and report an honest, attributable answer
rather than an unverified claim.
A Cyber Threat Intelligence Knowledge Graph for Exploit Link Prediction and Asset Risk Scoring
Existing CTI knowledge graph work stops at construction, extracting entities and relations from text but never reasoning over the finished graph. This study closes that gap by building a graph from structured feeds (MITRE ATT&CK, NVD, CISA KEV) and adding two reasoning layers: RotatE-based exploit link prediction, which roughly doubles the TransE baseline (MRR 0.173 vs 0.077), and an explainable graph-proximity risk score that ranks internal assets down to the CVE and actor behind each score. It also benchmarks LLM-based graph reconstruction against the same graph, showing structured feeds and text extraction are complementary. The core contribution is turning a static CTI graph into a tool that answers the questions analysts actually ask.
Personalized Thyroid Disorder Diagnosis System through Adaptive Expert Routing and Intelligent Selection
The main contribution of this work is AERIS, an adaptive expert-routing framework that performs patient-specific selection among heterogeneous machine learning models for thyroid disorder diagnosis. By combining a meta-learning router with Random Forest, LightGBM, CatBoost, and XGBoost, the framework achieves 97.87% accuracy and 95.98% F1-score, while SHAP-based analysis provides interpretable insights into the clinical features influencing predictions.
TriYield-XAI: An Explainable Ensemble model for Weather and Geospatial-Driven Aman, Aus, and Boro Yield Prediction in Bangladesh
We developed TriYield-XAI, a unified explainable framework for jointly predicting Aman, Aus, and Boro rice yields by integrating climatic, geographical, temporal, station-level, and historical yield information. We introduced a panel-aware hybrid ensemble architecture that combines Profile-KNN, matrix reconstruction, out-of-fold ensemble learning, meta-stacking, calibration, and residual correction to effectively capture complex spatio-temporal yield patterns and improve prediction performance. Furthermore, we applied SHAP-based explainable AI techniques to investigate the contribution of temporal, geographical, climatic, and station-specific factors to rice yield predictions, enhancing the interpretability and transparency of the proposed model.
A Fully Offline BLE-Connected Multi-Task TinyML Weather Station with Android Monitoring
The study’s main contribution is the integration of calibrated ultrasonic rainfall sensing, a three-output multi-task TinyML model on ESP32-S3, BLE communication, and an Android monitoring application into one fully offline weather station. This enables local next-day rainfall, humidity, and rainfall-risk prediction without relying on Wi-Fi, Internet, or cloud services.
Multi-Class Chest Disease Classification from X-Rays Using PCA-Enhanced Vision Transformers and Multi-Method Explainability
Main Research Contributions:
– Dimensionality reduction through PCA. To curb overfitting on uneven and heterogeneous data, we place Principal Component Analysis (PCA) inside the preprocessing stage, which lowers the number of features and the computational load while holding on to the information that separates the classes.
– Locality-aware attention transferred to radiography. We adopt the MLLKSA block of Luong et al., developed for breast histopathology, and fine-tune it on chest X-rays (CXRs). The contribution here is the transfer and its empirical assessment, not the mechanism itself.
– A transformer and CNN ensemble. To bring together their differing inductive biases, we join the fine-tuned Vision Transformer with convolutional networks in an ensemble setup. The transformer branch handles long-range global structure while the convolutional branch is stronger on local detail, and together they lift robustness and generalization.
– The ability to explain everything in detail. In order to guarantee that model decisions are clear, clinically reliable, and robust to shortcut learning, we incorporate a multi-method interpretability pipeline that uses gradient-based and model-agnostic visualization tools.
Design, Simulation, and Sizing Assessment of an Off-Grid Solar PV–PEM Green Hydrogen Generation and Compressed Storage System for Bangladesh
Existing PV–hydrogen studies often emphasize component modeling or techno-economic assessment, while integrated Bangladesh-specific frameworks remain limited. This study develops a MATLAB/Simulink-based solar PV–PEM system
using NASA POWER solar data, integrating PV generation, MPPT-controlled DC–DC conversion, PEM electrolysis, and compressed hydrogen storage. Mathematical verification is also performed for hydrogen production and storage performance.
Although LCOH is beyond the present scope, the results indicate strong potential for solar-driven hydrogen production in Bangladesh.
Do Macroeconomic Predictors Improve Bangladesh Inflation Forecasts? A Leakage-Safe Econometric and Machine-Learning Evaluation, 2010-2025 Abstract- This study evaluates whether lagged macroeconomic indicators improve
This study evaluates whether lagged macroeconomic indicators improve one-month-ahead forecasts of Bangladesh headline inflation beyond strong persistence-based benchmarks. A monthly data set covering January 2010 to December 2025 is assembled from public sources and contains year-on-year inflation, BDT/USD exchange-rate changes, the policy rate, Brent oil-price changes, and annual real GDP growth. To prevent information leakage, annual GDP is introduced only as the previous year’s realized value, and forecasts are evaluated through an expanding-window design on an untouched January 2023-December 2025 holdout. The analysis combines augmented Dickey-Fuller and KPSS stationarity tests, Ljung-Box diagnostics, a distributed-lag regression with heteroskedasticity-and-autocorrelation-consistent standard errors, Granger predictive-causality tests, SARIMAX, autoregression, Ridge regression, Random Forest, Extra Trees, Gradient Boosting, and XGBoost. Inflation is highly persistent, with a lag-one autocorrelation of 0.964. Exchange-rate lags are jointly significant in the explanatory model (Wald p < 0.001), whereas oil-price lags are not and policy-rate changes are only borderline significant. SARIMAX records the lowest RMSE (0.543 percentage points), but its improvement over the last-month benchmark is not statistically significant. Complex machine-learning models do not outperform the parsimonious benchmarks. The results indicate that exchange-rate information matters for inflation dynamics, but most short-horizon forecastability originates from inflation persistence. The paper demonstrates why chronological validation, strong naive baselines, and careful publication-timing controls are essential for credible macroeconomic machine learning.
Dynamic Performance Analysis of a 30kW PV System using P&O-Based Maximum Power Point Tracking
In this paper a detailed dynamic performance analysis of a 30-kW PV system using Perturb and Observe (P&O) maximum power point tracking (MPPT) algorithm is presented. As solar PV is increasingly integrated into today’s power networks, good MPPT is critical to harnessing maximum power. Although simple and inexpensive to implement, conventional P&O has a few significant disadvantages: oscillations around the maximum power point, loss of the direction of the power point tracking with fast changes in the solar irradiance, and compromising the direction of the tracking to speed up and increase the accuracy of the tracking. The proposed system is composed of PV array, DC-DC boost converter and P&O controller. The results of the simulations are compared and studied under the assumption of step changes in irradiance and load. The dynamic performance parameters like convergence time, oscillations under steady state condition, tracking efficiency and power quality are studied in detail. From the results presented, the performance of the standard P&O algorithm in steady state error tracking performance is good, while the performance in terms of dynamic response is limited, so there is a need to improve the performance of the dynamic response by using better variants of the P&O algorithm to achieve good performance in practice.
Detecting Financial Fraud Rings: Graph Motifs, Learned Detection, and a Base-Rate-Normalised Evaluation
1. Benchmark hygiene changes the answer. AMLworld’s final week has laundering at 59% purity against a 0.1% norm, holding 12.7% of all positives. A standard chronological split puts it in test. Removing it moves the graph-over-tabular advantage from 2.37× to 5.43×.
2. Precise graph queries degrade a GNN when injected as features. Detectors at up to 15.2× lift standalone cost 21% average precision and doubled seed variance as node attributes. Mechanism: they fire on 1.2% of accounts, so the columns are ~99% zero and dilute denser signal. This is a boundary condition on Blanuša et al., who report the opposite — their patterns attach to transactions and feed a tree ensemble, which can branch on a sparse indicator for free.
3. Two metadata columns beat 33 graph-derived ones. Ownership linkage — how many accounts share a legal owner — gave +10.7% AP and cut variance fourfold. Dense and semantic outperformed sparse and structural.
4. Base-rate-normalised lift, and the retention split. AP is uninterpretable across subsets of differing prevalence. Applied to leave-one-typology-out, recall retention is 93.5% but lift retention 45.5% — the model still finds an unseen coordination pattern, but ranks it far less confidently. Either figure alone misleads.
5. Empirical threshold selection for motif queries. Sweeping stored values rather than guessing more than doubled lift on three of four detectors — fan-out 6.2× → 15.2×, fan-in 2.7× → 12.9×
Academic Citation and Trend Forecaster: A Simple Graph Neural Network Approach for Venue Prediction and Collaboration Forecasting
A working end-to-end graph pipeline on a large real-world academic dataset — builds a heterogeneous graph (papers, authors, institutions) directly from OGBN-MAG’s raw files (736K+ papers, 5.4M+ citation edges), rather than a toy/small dataset.
A direct, controlled comparison of graph scale vs. architecture — full-graph GraphSAGE (mini-batch trained) vs. subgraph-sampled GAT, on the same task and dataset, showing that graph completeness (26.14% vs 9.24% accuracy) matters more than the choice of aggregation method (mean vs. attention). This is a useful, generalizable finding for anyone doing GNN work at scale.
A working link-prediction pipeline for collaboration forecasting — an encoder-decoder GraphSAGE model that scores author pairs, reaching 0.9998 ROC-AUC.
Methodological honesty as a contribution in itself — the paper explicitly flags that the near-perfect link-prediction AUC is inflated by random negative sampling and randomly-initialized author features, rather than presenting it uncritically. This kind of transparent limitation analysis is genuinely valuable and often missing from student/applied ML papers.
AI-Assisted Learning Tools in Improving Programming Skills Among University Students: A Case Study
This study investigates the impact of AI-assisted learning tools on programming skills among university students. It combines secondary analysis of an education dataset with a synthesis of recent computing-education research to examine changes in learning quality, learning time, stress, programming task performance, code comprehension, and debugging speed. The study also identifies the risk of over-reliance on AI tools and highlights the importance of structured pedagogical guidance for effective and responsible integration of AI in programming education.
Analysis and Prediction of Digital Distractions’ Impact on Academic Performance in Online Learning Using Machine Learning
This study combines survey-based behavioral analysis with supervised machine learning to predict the perceived academic impact of digital distractions and compares four classification models to identify the best-performing approach.
Integrating Deep CNN Models for Multilingual Sign Language Recognition: A SignLink-Based Approach for Bengali and English
Sign language is a vital medium of communication
for people who are deaf or hard of hearing. While much
research has addressed widely used systems such as American
Sign Language (ASL) and British Sign Language (BSL), Bengali
Sign Language (BdSL) remains far less explored. To help close
this gap, this work proposes a bilingual Bengali and English
sign language recognition system that applies deep learning for
stronger multilingual gesture interpretation. Two publicly avail
able datasets were merged. The first contains English letters from
A to Z, digits from 0 to 9, and a space symbol, while the second
holds 38 Bengali alphabet gestures. After the datasets were
combined and augmented to balance the class distribution, the
final collection held 75 classes and 112,493 images in total. Eight
deep learning models were evaluated, comprising six pre-trained
architectures and two custom networks. Among them, AlexNet
gave the best standalone test accuracy of 96.99%. To raise
accuracy further and reduce overfitting, a hybrid model named
SignLink was built by joining AlexNet with MobileNetV2 and
Xception, both of which generalize well. The hybrid reached a
test accuracy of 98.93%, surpassing every individual architecture.
The proposed system shows reliable bilingual recognition and
supports inclusive communication technology for linguistically
diverse and low resource settings.
VeriFake: A TF-IDF and Ensemble Learning Framework for Web-Based Fake News Detection
The rapid circulation of fabricated news has become a serious problem for society, since it spreads confusion,
reinforces bias, and erodes public confidence in trustworthy
information. Because digital platforms and social networks allow
misleading stories to travel far more quickly than before, they
often shape opinions, awareness, and everyday decisions. Reliable
and automated detection tools are therefore needed to limit
this damage. This study presents a supervised machine learning
framework that recognizes fake news directly from website text.
Term Frequency Inverse Document Frequency (TF-IDF) is used
for feature extraction, and thirteen classifiers are trained and
compared, namely Random Forest (RF), Decision Tree (DT),
Bernoulli Naive Bayes (BNB), Multinomial Naive Bayes (MNB),
Support Vector Machine (SVM), Linear Support Vector Classifier
(LSVC), K-Nearest Neighbor (KNN), Logistic Regression (LR),
XGBoost (XB), Multilayer Perceptron (MLP), Extra Trees (ET),
AdaBoost (AB), and Gradient Boosting (GB). The framework is
evaluated on a balanced Kaggle collection of 20,800 articles, split
evenly into 10,400 genuine and 10,400 false items. Random Forest
records the best accuracy at 99.23%, with LSVC and Decision
Tree following closely, while the remaining ensemble and boosting
methods also stay above 95%. For practical use, a secure Django
based web application was built so that users can submit news
text and receive an immediate prediction. The results confirm
that tree based ensembles and linear classifiers, when paired
with TF-IDF features, remain highly effective for monolingual
fake news detection, and the deployed platform shows how these
models can be turned into usable tools against misinformation.
Automated Code Smell Detection for Software Quality Assurance Using a Web-Based Machine Learning Framework
Code Smells are indicators of structural weaknesses
in software design and implementation that can reduce maintainability, scalability, and readability. To improve software quality
and support efficient maintenance, this study proposes a machine learning approach for the automated detection of Critical
Threshold Rule (CTR) violations, focusing on the Long Method
and Large Class smell types. In the feature selection stage,
six models, namely Decision Tree (DT), Random Forest (RF),
Logistic Regression (LR), Support Vector Machine (SVM), Multilayer Perceptron (MLP), and Stochastic Gradient Descent (SGD),
were combined with GridSearchCV, RandomizedSearchCV, Outof-Bag (OOB) validation, and SHAP values to raise both accuracy
and interpretability. After the best features were selected, thirteen
classifiers were trained and evaluated: LR, DT, RF, SVM,
Gaussian Naive Bayes (GNB), Multinomial Naive Bayes (MNB),
MLP, Linear Support Vector (Linear SV), K-Nearest Neighbors
(KNN), Gradient Boosting (GB), Extra Trees (ET), Bernoulli
Naive Bayes (BNB), and AdaBoost (AB). The models were trained
on datasets drawn from Software Development Versioning (SDV)
repositories that contain 1,000 log file entries. Across accuracy,
precision, recall, F1 score, and ROC AUC, ensemble models, and
RF in particular, gave the strongest results, reaching the highest
accuracy of 96.02% for Long Method and 92.63% for Large
Class. The findings were further confirmed by the Wilcoxon
signed-rank test. For practical use, the method was deployed
as a web application built with the React Native framework that
analyzes log data, flags smelly code segments, and reports their
severity to give developers actionable insight. The approach offers
a reliable and interpretable solution for Code Smell detection with
clear potential for integration into modern software development
End-to-End Explainable Stacked Ensemble Framework for Used Bike Price Prediction Using Real-World Marketplace Data
The significant research contribution of this study is the development of an explainable stacked ensemble framework for used-bike price prediction using real-world marketplace data from Bangladesh. The proposed framework integrates XGBoost, LightGBM, CatBoost, and HistGradientBoosting as base learners with Ridge regression as the meta-learner, supported by a comprehensive feature-engineering strategy that produces 28 predictors. Using 6,697 cleaned listings from Bikroy.com, the proposed model achieved an R^2 of 0.8759, RMSE of 0.2456, and MAE of 0.1712, outperforming all individually evaluated models. Furthermore, SHAP-based analysis provides interpretability by identifying engine capacity, vehicle age, and the brand–engine capacity interaction as major pricing factors. Finally, the complete framework was deployed as a 36 MB Android application using ONNX models, enabling on-device price prediction without cloud computation.
Pooled and Compromised: A Leakage Audit of a Popular MRI Benchmark
The Kaggle “Brain Tumor MRI Dataset” (Nick- parvar) is a saturated benchmark, with recent work reporting 97–99.9% accuracy under naive, patient-agnostic splits. It pools images from multiple sources without retaining patient identifiers, so near-duplicate slices from the same scan can land on both sides of a random split, inflating accuracy. We attempt to recover true patient identity for the Figshare-derived portion (glioma, meningioma, pituitary) by matching 7,200 pooled Kaggle images against the original 3,064 patient-labeled Figshare images, using a two-stage pipeline: normalized perceptual-hash retrieval with SSIM confirmation, then an ORB feature-matching fallback. This recovers a patient ID for 527 of 5,400 eligible images (9.8% coverage), a 49-fold improvement over a single-stage baseline (0.2%), though coverage remains far below what the dataset’s documented provenance implies, and only 52.2% class- label agreement among recovered images suggests many ORB- fallback matches are incorrect. We combine these identities with near-duplicate clustering over the remaining pool (24.7% multi- image clusters) to build a group-aware, leakage-corrected split covering all 7,200 images. Retraining three CNNs (ResNet50, EfficientNet-B0, DenseNet121) under both protocols shows the naive split overstates test accuracy by 0.93–1.59 points per architecture. A soft-voting ensemble reaches 98.06% accuracy under the corrected protocol, significantly better than the best single architecture, EfficientNet-B0, at 97.26% (McNemar’s test, p = 0.012), reversing a null ensembling result from a weaker, purely clustering-driven correction in an earlier pass. We release the corrected split, code, and all figures/tables for reuse.
Impact of Organic Fertilizer on Yield Soil and Health
Evaluation of Sustainable Nutrient Management: This paper provides a comprehensive review of structural, chemical, and biological transformations in soil health driven by organic fertilizers, biochar, cover crops, and bio-amendments such as Plant Growth-Promoting Rhizobacteria (PGPR) to counter intensive chemical fertilization.
Quantitative Analysis of Soil Properties: It synthesizes empirical datasets to deliver a quantitative breakdown of soil physical attributes (macro-aggregate stability, bulk density, porosity), chemical parameters (pH, electrical conductivity, sodium adsorption ratio, and soil organic matter), and biological/enzymatic activities like urease and phosphatase.
Crop Productivity and Quality Assessment: The research investigates physiological responses including chlorophyll accumulation (SPAD index), proline stress alleviation, and ultimate yield quality improvements such as protein, gluten, and total soluble solids across multiple crop species.
Integration of Advanced Technologies and Challenges: It discusses practical application constraints, nutrient release kinetics, and future technological integrations including GIS, Remote Sensing, and Precision Agriculture to optimize organic fertilizer management.
Bridging the Digital Divide in Secondary Education in Rural Bangladesh: Challenges, Impacts, and Policy Solutions.
This research provides a systematic review of the digital divide in secondary education in rural Bangladesh by analyzing findings from 45 studies. It brings together the major challenges, including poor internet infrastructure, socio-economic inequality, gender gaps, and limited digital skills. The study also examines how these issues affect students’ learning and identifies gaps between existing digital education policies and their implementation. Finally, it suggests practical directions for improving digital access, teacher capacity, and educational opportunities for rural students.
Retrospective Landmark Classification of Recorded Administrative Student Dropout Status from University Records
This study advances the methodology of educational data mining by establishing a leakage-controlled retrospective landmark evaluation framework that exposes the analytical boundaries of predicting student attrition from administrative records. Rather than treating repeated final administrative categories as dated withdrawal events, this research demonstrates why high random-split discrimination should not be mistaken for prospective intervention readiness.
A Comparative Analysis of COVID-19 Vaccine Side Effects vs. Environmental Allergies: A Machine Learning Approach to Symptom Differentiation
This research contributes a machine learning-based approach to distinguish COVID-19 vaccine-related side effects from environmental allergy symptoms. By analyzing demographic, vaccination, symptom, and environmental exposure factors, the study identifies patterns that may help improve symptom classification and support more informed clinical assessment. The research also provides a foundation for developing data-driven tools for differentiating vaccine-related reactions from common environmental allergies.
Machine Learning Techniques in Cybersecurity Threat Detection: A Systematic Literature Review
There is a systematic analysis of machine learning techniques for cybersecurity threat detection provided in this study. It contrasts conventional machine learning, deep learning, and ensemble methods in the most prominent categories of cyber threats, and reveals important challenges like adversarial attacks, imbalanced data sets, model interpretability, concept drift, and limited ability to generalize. Another key finding of the study is that it identifies several research gaps and opportunities that would guide future research in more adaptive, explainable, and region-specific cybersecurity solutions.
The Impact of Generative Pre-trained Transformer and Large Language Models on Students’ Academic Performance
This study combines machine learning and statistical analysis to identify key LLM usage behaviors associated with students’ perceived academic improvement. Random Forest achieved 82.6% accuracy and an F1-score of 0.895, while the findings highlight the importance of responsible and critical LLM use.
An Analysis Between Public and Private Universities
This review synthesizes existing evidence on inequalities between public and private higher education institutions. It examines disparities in funding, affordability, educational quality, infrastructure, student satisfaction, and graduate employability. The study develops a comparative understanding of these institutional differences while identifying key methodological and empirical gaps in the existing literature. The findings provide a basis for informed policy development and future research toward greater equity in higher education.
An Investigative Survey on Depression Detection from Social Media Data Using Machine Learning and Deep Learning Models
This study reviews different machine learning and deep learning methods used to detect depression from social media data. It compares commonly used models and techniques and discusses their strengths, limitations, and challenges. The study also highlights possible directions for improving the accuracy and reliability of depression detection systems.
Kidney Disease Detection Using Machine Learning with Explainable Artificial Intelligence
This research provides a comprehensive systematic review of machine learning-based approaches for Chronic Kidney Disease (CKD) detection with a focus on Explainable Artificial Intelligence (XAI). The study identifies the effectiveness of widely used ML models, including Random Forest and XGBoost, and analyzes how explanation techniques such as SHAP and LIME improve model transparency. Additionally, it highlights key research challenges, including limited dataset diversity, insufficient clinical validation, and the need for interpretable AI solutions in healthcare. The findings provide valuable insights for developing reliable, transparent, and clinically applicable CKD detection systems.
Urban and Rural Areas Studies Difference Using Machine Learning
This research contributes a machine learning-based approach for identifying and analyzing the differences between urban and rural areas using socioeconomic and demographic features. The study compares multiple machine learning models and demonstrates that Random Forest can effectively classify urban and rural areas, achieving an accuracy of 94.20%. The findings provide data-driven insights that can support better planning, resource allocation and strategies for reducing urban–rural disparities.
Impact of Social Media on Mental Health
This research study that how affecting on our mental health in every possible way. we identifies some risk factors and highlights how machine learning and deep learning techniques can detect mental heatlh issue from social media data.
Educational Differences Between Urban and Rural Schools: Challenges and Opportunities in Rural and Urban Education
This study looks at the differences between rural and urban education in Bangladesh using survey data from 100 students. It focuses on factors such as teachers, internet access, transportation, and school facilities. Three machine learning models were also tested, with Random Forest giving the best results.
Impact of social media on students academic performance
This study examines how social media use affects students’ academic performance, study habits, and concentration, providing insights for promoting balanced and effective social media use.
Predicting Dinosaur Coloration from Fossilized Melanosome Data Using Classical Machine Learning
This study establishes a reproducible classical machine-learning baseline for dinosaur coloration reconstruction from fossilized melanosomes, while systematically evaluating cross-clade generalization. By comparing QDA, SVM, and Random Forest on theropod and non-theropod taxa, the study demonstrates that high-confidence predictions for phylogenetically distant specimens reflect distribution shift and morphological overlap with the avian training data rather than reliable biological coloration. These findings highlight the limitations of transferring bird-trained models across dinosaur clades and motivate more taxonomically broad and taphonomically informed approaches to paleocolor reconstruction
Explainable Machine Learning framework for Antibiotic Resistance Prediction in E. coli UTIs: Clinical Insights for Bangladesh
Urinary tract infections caused by Escherichia coli are among the most common bacterial infections in clinical practice. Increasing antibiotic resistance has made empirical treatment increasingly challenging particularly in resource-limited settings such as Bangladesh. This study proposes an explainable machine learning framework for predicting antibiotic resistance in E. coli UTI patients using routine clinical and laboratory data. Two publicly available datasets were merged to create a
dataset of 10,710 patient records with 32 processed features. Five machine learning models including Logistic Regression, Decision Tree, Random Forest, Support Vector Machine and XGBoost were evaluated using 5-fold stratified cross-validation across five antibiotic resistance targets including Ciprofloxacin, Gentamicin, Amoxicillin-Clavulanic Acid, Cefotaxime/Ceftriaxone and Amoxicillin/Ampicillin. SHAP and LIME were employed to provide global and patient level model explanations. SVM achieved the best overall performance for AMC, CTX/CRO and AMX/AMP with the highest AUC of 0.586 for CTX/CRO, while XGBoost obtained the highest accuracy for CIP (83.87%) and GEN (78.33%) but detected almost no resistant cases at those operating points. SHAP identified age, diabetes status, body temperature, hospital admission history, hypertension and the Neutrophil-to Lymphocyte Ratio as the most influential predictors, consistent with risk factors reported in Bangladeshi clinical studies. These findings demonstrate that explainable machine learning using routinely available clinical data can support early antibiotic resistance prediction and assist empirical antibiotic selection in resource-limited healthcare settings.
Predicting E-Learning Efficacy and Gap Closure in Post-Pandemic Higher Education Using Machine Learning
The COVID-19 pandemic shifted higher education from traditional in-person classrooms toward flexible digital delivery. While emergency remote teaching preserved instructional continuity during lockdowns, it also widened academic disparities. This study examines whether online education has since closed the learning gap it helped create. We analyzed 509 survey responses collected across the eight administrative divisions of Bangladesh, yielding 284 unique response profiles after duplicate removal. Using a quantitative supervised machine learning framework, we evaluated seven socio-technical predictors: administrative division, internet access quality, device availability, teacher online class hours, student engagement, community readiness, and institutional support. Random Forest, K-Nearest Neighbors, and Support Vector Machine classifiers were tuned by grid search with 5-fold cross-validation on the training partition and assessed on a held-out test set. Random Forest achieved the strongest holdout performance, with a test accuracy of 84.21% and a weighted F1-score of 0.83, while SVM attained the highest cross-validation accuracy at 83.29%. Feature importance analysis identified Institutional Support as the dominant contributor at 0.54, followed by regional context at 0.12. Across the evaluated classifiers, the surveyed cohort predominantly reflected a perception that the educational gap was only partially closed. These findings indicate that although online platforms have improved flexibility, access, and instructional continuity, virtual tools alone cannot replicate the collaborative presence of physical classrooms. Digital learning therefore works best as a complementary component of a balanced hybrid model rather than a full replacement for traditional instruction.
A Machine Learning-Based Comparative Analysis of Rural and Urban Educational Inequality in Bangladesh
Educational inequality between rural and urban areas remains a significant challenge in Bangladesh, where differences in infrastructure, technology, and academic support shape students’ learning outcomes. This study investigates these disparities using survey data collected from 145 respondents, organized into separate primary and high school datasets and analyzed using descriptive statistics and supervised machine learning. The analysis focuses on educational resources, digital facilities, internet access, private tutoring, teacher support, and academic performance. Four classification models were com pared: Logistic Regression, Support Vector Machine (SVM), Random Forest, and Extreme Gradient Boosting (XGBoost), each trained to classify rural and urban educational backgrounds and evaluated using accuracy, precision, recall, F1-score, and five fold cross-validation. The results reveal substantial rural–urban differences in educational resources, with urban students reporting higher internet availability, greater access to digital devices, more computer laboratory facilities, and higher rates of private tutoring, and better facilities associated with improved academic performance. Logistic Regression achieved the highest accuracy, reaching 79% for the primary school dataset and 83% for the high school dataset. The findings highlight greater inequality at the high school level and demonstrate the potential of machine learning to identify educational disparities, supporting the need for improved educational infrastructure and digital accessibility in Bangladesh.
Mental Health Classification from Social Media Data in Banglish using Transformer-Based Multi-Task Learning and Explainable AI
1. We have released a public Banglish mental health dataset of 56,522 samples.
2. We have presented a multi-task architecture named BanglaBERT-MTL-AP.
3. We have introduced a three-layer hierarchy routing module and Grad*Input based XAI system for the interpretability.
IETS-Stroke: An Imbalance-Aware, Explainable, and Threshold-Tuned Stacking Framework for Stroke Risk Screening with Cross-Dataset Validation
This research proposes IETS-Stroke, an imbalance-aware, explainable, and threshold-tuned stacking framework for stroke risk screening using public-health datasets. The framework combines imbalance handling, stacking ensemble learning, threshold optimisation, SHAP explainability, and BRFSS-to-NHANES external validation. It can be used as a screening-support approach to identify individuals who may require further clinical assessment, rather than as a diagnostic system.
Confidence-Aware Retrieval-Augmented Generation with Abstention for Cyber Threat Intelligence
Retrieval-Augmented Generation (RAG) grounds
language models in external knowledge, which makes it a
natural tool for Cyber Threat Intelligence (CTI). It has one
dangerous habit, however: it answers every question, even when
the retrieved evidence does not support an answer. The result
is hallucinated vulnerability details that can mislead a security
analyst. We present a confidence-aware RAG framework for
retrieving CVE-based threat intelligence. The framework adds a
tunable abstention step. When its confidence in the retrieval is too
low, the system declines to answer instead of guessing. We measure
this confidence at the retrieval stage from two simple signals: how
similar the best-matching record is to the query, and how far
ahead it is of the second-best record. The measure needs no access
to the language model’s internals. We evaluate three systems,
Naive RAG, Hybrid RAG, and our Confidence-Aware RAG, on
200 keyword-based queries over the 2023 National Vulnerability
Database (30,932 CVE records). At a confidence threshold of
0.20, our system answers with 81.3% precision, against 70.0%
for a baseline that always answers, and it lowers the Expected
Calibration Error from 0.309 to 0.259. In short, for security-
critical CTI work, a few trustworthy answers are worth more
than many unreliable ones.
Calibrated and Explainable Credit Risk Assessment: Integrating SHAP Analysis with Feature Stability Validation
Machine learning has substantially improved the predictive performance of credit risk assessment, yet high pre- dictive accuracy alone is insufficient for deployment in regu- lated lending, where reliable probability estimates, interpretable decisions, and consistent explanations are equally important. Existing studies typically evaluate these properties independently, leaving uncertainty about whether a model can satisfy all three simultaneously. This paper proposes an integrated evaluation framework that jointly assesses discrimination, probability cali- bration, and explanation stability with statistical confidence. Us- ing the German Credit dataset, we compare Logistic Regression, Random Forest, and LightGBM. Model performance is evaluated using ROC-AUC, the Brier score, Expected Calibration Error (ECE), bootstrap confidence intervals, and paired significance tests, while TreeSHAP explanations are validated across five cross-validation folds using feature-wise coefficients of variation. Experimental results show that LightGBM achieves the best overall discrimination and calibration while maintaining stable explanations for its most influential risk factors. However, its performance advantage over Random Forest is not statistically significant. We further demonstrate that synthetic oversampling improves recall but consistently degrades probability calibration, whereas cost-sensitive threshold optimization achieves compara- ble gains without distorting predicted probabilities. In addition, a fairness audit reveals demographic disparities that are not reflected in feature attribution alone. These findings highlight that discrimination, calibration, explanation stability, and fairness should be evaluated together, providing a more reliable basis for selecting and validating credit risk models for practical deployment. Index Ter
Agentic AI in Software Engineering: Evidence from Production Software and Future Directions for Developing Countries
The significant research contribution is the empirical evaluation of agentic AI coding tools in a production-style software environment. The study uses 160 tasks performed by 16 professional .NET developers on the nopCommerce platform to quantify success, task-complexity effects, human intervention, token exhaustion, and failure patterns. It demonstrates that task complexity significantly affects agentic AI performance and provides practical recommendations for human-in-the-loop adoption of agentic AI in Bangladesh and similar developing-country software environments
Retrieval-Augmented Legal Question Answering Using Hybrid Semantic Search
1. A full ablation of sparse, dense, fusion-based and document-aware retrieval on the complete benchmark, scored with a character-level metric.
2. A weighted fusion scheme that repairs the failure of equal-weight reciprocal rank fusion, and a sparse backend (bm25s) that cuts BM25 scoring time by three orders of
magnitude.
3. A two-stage document-aware retriever that narrows the search to the correct document before re-ranking, the strongest method we tested.
4. A citation-grounded generation stage whose citations are scored against gold spans, showing that better retrieval yields better-attributed answers even though the model tends to under-cite.
A Multi-Dimensional Performance Assessment of Rooftop Agrivoltaic Systems for Sustainable Urban Development
This research contributes to the development of agrivoltaic systems by investigating the dual-use of agricultural land for photovoltaic electricity generation and crop cultivation, with emphasis on optimizing PV installation conditions to maintain energy production while supporting agricultural productivity. The findings provide practical insights for sustainable land use, renewable energy generation, and climate-resilient agriculture, particularly in regions such as Bangladesh.
Experimental Investigation of Rooftop Solar PV Deployment for Passive Building Cooling and Sustainable Electricity Generation
This study demonstrates the dual functionality of rooftop photovoltaic (PV) systems as both renewable energy generators and passive thermal-shielding elements. The key contribution is the experimental quantification of rooftop cooling using thermal imaging across three different PV installation heights, showing an average 25.97 °C (46.14%) reduction in rooftop surface temperature. The findings provide practical evidence that strategic PV deployment can simultaneously reduce building heat gain, potentially lower cooling demand, and support clean energy generation, offering a practical approach for energy-efficient and climate-resilient urban buildings.
The Adoption of Internet of Things in Higher Education: Opportunities, Challenges, the Role of vision 2030 in Saudi Arabia
Empirical baseline — Establishes that 65% of Saudi universities have deployed some IoT, with adoption skewed toward well-resourced institutions, giving policymakers a concrete starting point.
Integrated theoretical lens — Applies TAM and DOI together to explain why adoption succeeds or stalls in this specific national context, not just whether it does.
Three-factor barrier model—Identifies and categorizes the interplay of technical, economic, and cultural-organizational barriers as a coherent obstacle framework specific to Saudi Arabia, rather than treating each in isolation.
Vision 2030 alignment — Positions IoT as a measurable instrument of national policy, connecting technology adoption research to a live strategic agenda—giving the findings direct policy relevance beyond academic circles.
Actionable recommendations—Translates findings into three concrete policy levers: infrastructure investment, training programs, and regulatory frameworks.
Explainable Multi-Class Electricity Theft Detection on the TDD2022 Benchmark
This work presents the first explainable, per-attack-type analysis of electricity theft detection on the TDD2022 benchmark. Ten classifiers are evaluated under a frozen, temporally ordered, per-building-type split that prevents leakage between training and test sets. Rather than reporting aggregate accuracy alone, the study reveals that detection difficulty varies dramatically by attack type from perfect detection to near-complete evasion and uses SHAP to attribute why each attack is caught or missed to specific consumption features. A McNemar’s test across ten models, including modern gradient-boosting methods, demonstrates that performance is bounded by the feature representation rather than model choice, motivating sequence-aware detection as future work.
A Human-Centred Critical AI Literacy Framework for Higher Education
This research contributes a novel Seven-Stage Critical AI Literacy Framework that shifts higher education responses to GenAI from surveillance and compliance towards the development of evaluative judgement, ethical reasoning and reflective accountability. The framework provides a structured, assessment-oriented approach for critically evaluating AI-generated outputs across seven dimensions and makes human judgement and intellectual accountability visible and assessable. It offers a practical foundation for assessment redesign, curriculum development and responsible AI integration across disciplines.
Real-Time Food Spoilage Detection Using Multi-Sensor Gas Analytics and IoT-Based Monitoring
This study introduced an IoT-enabled real-time food rotting detection architecture powered by dynamic edge computing and multi-sensor gas analytics. The system effectively adjusted its categorization limits (Td) to real-time microclimate fluctuations by merging MQ-series gas data into a single index (ΦGas) and using a continuous multivariate contextual compensation function. The system successfully neutralizes baseline drift noise caused by environmental humidity and temperature fluctuations while achieving an absolute classification accuracy of 100% across transitional decomposition phases, according to experimental validation using telemetry streams.
Should Bangladesh Adopt Age Verification Laws? A Socio-Legal Analysis of Surveillance, Privacy, and Public Trust
This study examines the implications of age-verification (AV) laws for Bangladesh, where digital governance is evolving but public trust in data handling remains low.
The paper recommends rights-based alternatives, including digital literacy initiatives, parental empowerment, and privacy-preserving age assurance, as more context-appropriate strategies for protecting children online.
A Dual-Backbone Hybrid Deep Learning Framework with Squeeze-and-Excitation Attention and Grad-CAM Explainability for Multi-Class Brain Tumor MRI Classification
Accurate brain tumor classification from magnetic
resonance imaging (MRI) is essential for treatment planning,
yet manual interpretation remains time-consuming and subject to inter-observer variability. This paper proposes a dual-
backbone convolutional neural network that fuses EfficientNetB3
and MobileNetV2 features through a Squeeze-and-Excitation
(SE) channel-attention block for classifying glioma, meningioma,
pituitary tumor, and no-tumor MRI scans. The preprocessing
pipeline combines Gaussian–bilateral denoising with contrast-
limited adaptive histogram equalization (CLAHE) to suppress
noise while preserving diagnostically relevant contrast. The model
was trained progressively through frozen-backbone warm-up,
partial fine-tuning, and full end-to-end optimization with cosine
learning-rate scheduling on 7,153 MRI images. The proposed
network achieved 99.72% validation accuracy and 99.16% test
accuracy, with macro precision, recall, F1-score, and ROC-AUC
values of 0.9914, 0.9917, 0.9915, and 0.9997, respectively. Ablation
analysis showed that direct backbone fusion without SE attention
reached only 95.35% accuracy, whereas the complete attention-
enhanced model achieved 99.16%, confirming the importance
of adaptive channel recalibration. Grad-CAM was used to
visualize class-relevant tumor regions and improve prediction
transparency. These findings demonstrate that the proposed
attention-guided dual-backbone framework provides accurate,
interpretable, and computationally practical brain tumor classification for clinical decision-support applications.
A Cross-Environment Benchmarking Study of Quantum Machine Learning for Wi-Fi Deauthentication Attack Detection Using Public Datasets and Real-World Captures
This study provides a leakage-resistant, cross-environment comparison of quantum and classical models for Wi-Fi deauthentication detection using public and real-world data, including target-domain adaptation and actual quantum hardware evaluation.
Real-Time Object Classification in Autonomous Driving: A Lightweight Residual CNN Approach on Unified BDD100K–COCO Data
Unified Multi-Dataset Super-Taxonomy: A novel three-class (Vehicle, Pedestrian, Obstacle) crop-based data unification pipeline integrating BDD100K and MS COCO 2017, yielding 58,831 annotated ROI samples with a consistent label space for real-time classification.
Lightweight Residual CNN Architecture: A custom four-stage residual backbone (channels 32→512) with global average pooling and dropout regularisation, specifically designed for resource-constrained on-board automotive hardware.
Speed–Accuracy Optimisation: Achievement of 215.5 FPS inference throughput via 224×224 input resolution and FP16 mixed-precision acceleration while maintaining 93% mean classification accuracy and per-class F1-scores of 0.94/0.94/0.91.