Hezerul Abdul Karim

Introduction: Feature selection helps identify important features in high-dimensional mental health data. However, conventional feature selection (CFS) methods commonly rank all features within a single feature space using an overall statistical or predictive criterion. This global evaluation method does not take into account the fact that symptoms of depression are interrelated and fall within various distinct clinical domains. To address this limitation, we propose partitioned feature selection (PFS), a framework that integrates domain-based feature grouping with group-wise feature selection. Methods: PFS organizes depression symptoms into clinically relevant domains and applies four complementary feature selection methods to estimate individual feature importance within each domain. The importance scores are normalized and combined, then weighted according to each domain's association with global severity to calculate integrated group feature importance (IGFI). We evaluated PFS on two independent multiclass depression datasets using 16 classifiers. Validation-guided feature retention was used to select the final feature subsets. We also conducted a ten-trial random regrouping analysis, an exploratory Wilcoxon signed-rank comparison with CFS-RFE, and a comparison of IGFI rankings with SHAP explanations. Results: PFS selected 21 features for Dataset 1 and 23 features for Dataset 2. Using the selected, untuned support vector classifier (SVC) configuration without synthetic minority over-sampling technique (SMOTE), PFS achieved test accuracies of 95.65% and 89.60% and Macro-F1 scores of 95.43% and 92.37%, respectively. The random regrouping analysis showed that the predictive effect of clinical partitioning varied between datasets. The exploratory Wilcoxon analysis also showed dataset-dependent statistical differences between PFS and CFS-RFE. Furthermore, IGFI and SHAP identified similar important features across clinical domains. Discussion: PFS provides a structured approach to identifying important depression symptoms while preserving their clinical domain information. The findings indicate that domain-based grouping can support interpretability, although its predictive benefits vary across datasets. This study focuses on questionnaire-based depression severity estimation rather than the prediction of an independently established clinical diagnosis.

Deep learning fire detectors are usually optimised for in-domain accuracy, while their cross-dataset generalisation, which is vital for real-world deployment, remains poorly understood. We propose a dual-teacher knowledge distillation framework that compresses a YOLOv8m teacher (25.84M parameters, 78.7 GFLOPs) into a GhostConv-enhanced YOLOv8n student with efficient channel attention (2.298M parameters, 6.2 GFLOPs) through spatial feature mimicry and CLIP ViT-B/32 semantic guidance. The distilled student achieves 72.0\% mAP@50 on the D-Fire validation set (four-seed mean), retaining 91.6\% of teacher performance with 11.2$\times$ fewer parameters, and 71.5\% mAP@50 on the held-out test set at 375.6 FPS with TensorRT on an RTX 5080, exceeding the closest distillation baseline by 9.3 percentage points. Ablations show that CLIP semantic distillation alone reaches 72.6\% mAP@50, surpassing the combined loss and revealing the importance of the transfer mechanism under architectural mismatch. Across four external benchmarks, however, all eight tested models fall from 71--79\% mAP@50 on D-Fire to $\leq$3.3\%, and CLIP-based confounder prompts do not reduce false positives relative to detection-only training. This first multi-benchmark generalisation analysis of a sub-5M-parameter fire detector shows that distillation delivers edge-deployment gains but leaves cross-domain robustness unresolved.

Depression is a highly prevalent mental health condition worldwide, affecting approximately 332 million individuals and contributing significantly to global disability burden. In computational mental health research, machine learning approaches are increasingly applied to structured psychometric questionnaires; however, many existing studies formulate the problem as binary classification, which fails to capture the ordinal nature of symptom severity represented in standardized instruments. This study presents a computational psychometrics framework for depression severity stratification using the Depression Anxiety Stress Scales (DASS-42) questionnaire dataset. The proposed model, DepressNet, four base learners—XGBoost, Random Forest, LightGBM and a calibrated support vector machine (SVM)—with a PyTorch-based multi-layer perceptron meta-learner trained on out-of-fold probability features to reduce information leakage and improve generalization. To address class imbalance inherent in real-world psychometric data, the training pipeline incorporates SMOTETomek resampling together with class-weighted cross-entropy loss. The framework is evaluated on 39,775 real-world questionnaire responses under an imbalanced five-class classification setting using multiple complementary metrics. On a held-out test set, the model achieves 85.38% accuracy, 59.01% macro F1-score, 0.7158 Matthews Correlation Coefficient, and 0.9714 AUC-ROC. It is acknowledged that the severity labels are derived from the same DASS-42 items used as input features; consequently, the framework learns structured scoring patterns within psychometric data rather than predicting independent clinical ground truth. The Mild severity class remains the most challenging stratum (F1 = 0.10), reflecting inherent symptom overlap with adjacent classes and the persistent difficulty of classifying a narrow four-point severity band. Beyond predictive evaluation, a SHAP-based explainability module is applied to examine both global and class-wise feature attribution patterns across severity levels. The analysis provides interpretable relationships between questionnaire items and predicted severity categories, highlighting structured symptom-response patterns captured by the model. Overall, this work contributes a reproducible and interpretable machine learning framework for psychometric data analysis, with emphasis on severity stratification, ensemble-based learning, and explainable modeling of mental health questionnaire responses.

Complex environment variability and limited computational resources in agricultural settings create challenges in automated detection in tropical regions, where jackfruit diseases pose major threats to harvest productivity and food security. Conventional machine learning models often fail to generalize across diverse image conditions. For sustainable, field-deployable diagnostic systems, it is essential to close the performance–efficiency gap, as traditional deep learning methods require large amounts of training data and high-end hardware. We propose a hybrid deep learning machine learning (DL-ML) framework that has been tailored for accurate and resource-efficient jackfruit disease identification. This framework discusses deep visual representations that are extracted using a lightweight custom Convolutional Neural Network (CNN) combined with a previously trained EfficientNet-B0, capturing both localized and nested image features. Principal Component Analysis (PCA) is utilized for feature optimization to enhance discriminative power and eliminate redundancy, while SMOTE minimizes dataset imbalance by generating synthetic minority samples. These selected features are then classified by employing a variety of machine learning algorithms, including Random Forest (RF), Support Vector Machine (SVM), XGBoost, CatBoost, and k-Nearest Neighbors (KNN). A stacked ensemble model incorporates the strengths of these classifiers, enhancing generalization and robustness under diverse lighting and background conditions. Experiments across multiple dataset configurations show a noticeable improvement in accuracy from 76.80% on raw data to 99.39% on the PCA-augmented dataset, illustrating that this hybrid ensemble consistently outperforms all individual deep and machine learning baselines. Although the approach is currently tailored to jackfruit disease datasets, its applicability to other crops or stress conditions requires greater domain adaptation studies. However, the model's high accuracy highlights its scalability and utility in low-resource agricultural environments. Additionally, it establishes a cost-effective diagnostic paradigm where efficient ML classification and deep feature extraction collaborate, laying the foundation for next-generation smart farming and precision agriculture systems.

Traffic sign recognition is a critical aspect of intelligent transportation systems that enhances road safety, traffic management, and driver assistance. Notably, while significant research efforts have been devoted to many prevalent traffic signs, Bangladeshi sign recognition remains relatively underexplored due to a lack of traffic sign knowledge and unawareness of the sector individuals. The scarcity of Bangladeshi traffic sign datasets has exacerbated this limitation, hindering progress. Moreover, most Bangladeshi traffic sign research predominantly relies on a limited number of signs, leading to diminished accuracy and the inability to detect all traffic signs. Additionally, existing systems grapple with suboptimal performance accuracy and heightened computational complexity, further emphasizing the research gap. We propose a robust Bangladeshi traffic sign recognition system to address these formidable challenges using a deep learning-based Convolutional Neural Network (CNN) architecture. The main aim is to create a lightweight model that helps identify traffic signs correctly with less computational effort. We used the effective combination of various deep learning layers in the proposed system. A significant contribution of our work lies in creating a novel Bangladesh Traffic Sign Recognition (BDTSR) dataset, addressing the scarcity of data in this domain. This dataset comprises comprehensive information, including 48 types of traffic signs, and these images were taken from different roads in the country. Our dataset aims to fill a critical gap in Bangladeshi traffic sign research and provides a solid foundation for more extensive and inclusive studies in the field. Through this innovative approach, we aim to contribute significantly to the field of BDTSR, filling the gaps in traffic sign recognition and bolstering the accessibility of traffic signs within the Bangladeshi individual traffic community and beyond. Our evaluation of the newly created BDTSR dataset and the existing two benchmark datasets resulted in high recognition accuracies of 99.34%, 96.97% and 98.82% accuracies for the BDTSR, German Traffic Sign Recognition Benchmark (GTSRB) and the Belgian Traffic Sign Dataset (BTSC) respectively. These results highlight the superiority of our model in the Bangladeshi Traffic Sign Recognition domain, outperforming existing models in terms of accuracy and computational efficiency.

Human activity recognition (HAR) in low-light conditions remains challenging due to photometric domain shift, reduced contrast, altered colour statistics, and noise patterns absent from standard RGB pre-training data, causing models to degrade even after fine-tuning. To address this, we propose a lightweight adaptation framework leveraging frozen CLIP representations that ensures computational efficiency while delivering competitive performance across the three evaluated benchmarks. Our approach extracts embeddings from uniformly sampled frames without backbone updates, avoiding the computational overhead of full fine-tuning. By employing lightweight parameter-efficient adapters, we enable effective adaptation with minimal computational overhead. This design yields measurable accuracy improvements under low-light photometric shift. For instance, temporal adapters with only 88.8K parameters improve ARID v1.5 accuracy from 68.37 ± 0.49% to 74.38 ± 0.13%, an absolute gain of 6.01 percentage points. We further demonstrate the benefits of this approach in other settings: (1) linear probing achieves 87.26 ± 0.50% accuracy on UCF101 at 17.44 GFLOPs, within 6% of fine-tuned SlowOnly (93.16% at 27.3 GFLOPs) with 36% fewer computations, highlighting frozen CLIP’s efficiency as a baseline for standard datasets; and (2) zero-shot classification shows a performance gap of approximately 15–46 percentage points across datasets, underscoring the necessity of task-specific adaptation. All lightweight adapters operate at essentially the same cost as the frozen backbone since temporal adapters, LoRA, attention pooling, and prompt tuning each add under 0.2% additional FLOPs. These results establish frozen vision-language models with lightweight adaptation as a scalable and efficient framework for HAR under photometric domain shift, demonstrated on low-light infrared and standard RGB benchmarks, achieving meaningful gains with no backbone retraining required.

Introduction: The reliability of phishing Uniform Resource Locator (URL) detectors under adversarial URL rewriting, domain shift, and explanation instability remains insufficiently understood. This study proposes an integrated robustness evaluation protocol for URL-based phishing detection, which integrates structured adversarial perturbation, unseen attack-family generalization, compositional attack effects, explanation stability, and external vulnerability transfer. Methods: The protocol tests four representative model families: Logistic Regression, XGBoost, CharCNN, and BERT-base, using 235,370 validated URLs from PHIUSIIL, consisting of 100,520 phishing and 134,850 benign URLs, along with 49,121 PhishTank-validated phishing URLs for external validation. Results: All models performed well on the clean test sets, ranging from 0.9962 to 0.9984, but their robustness decreased substantially under realistic URL mutations. Subdomain injection degraded the strong performance of Logistic Regression, XGBoost, and CharCNN to around 0.432, indicating collapse to the phishing-prevalence floor. BERT was highly susceptible to homoglyph, padding, and path-based perturbations. Leave-one-family-out evaluation also showed poor transfer to unseen subdomain attacks for both Logistic Regression and XGBoost, with Robustness Degradation Index values of 0.535 and 0.565, respectively. Explanation stability also suffered, with SHAP top-K Jaccard similarity dropping to 0.526-0.535 under subdomain perturbation. Discussion: These results provide a solid benchmark for evaluating robustness-aware phishing URL detection for achieving deployable reliability under realistic adversarial and non-IID settings.

Artificial intelligence systems are often assumed to become fairer when training datasets include broader demographic representation. This article argues that such a view is insufficient, especially when AI models developed in high-resource settings are exported to the Global South. Drawing on evidence from healthcare, biometric systems, natural language processing, public-sector algorithms, and financial technologies, the paper shows that inequity emerges across multiple stages of the AI pipeline: biased data construction, aggregate evaluation metrics that conceal subgroup harms, missing post-deployment feedback loops, and weak generalization under distributional shift. These failures are not isolated technical issues but interconnected structural problems that can amplify existing global inequalities. The article proposes that representation should be understood not as a demographic checklist, but as infrastructure: a system of data practices, contextual validation, epistemic inclusion, monitoring mechanisms, appeal pathways, and governance capacity. Without these conditions, AI systems risk scaling inequity while appearing technically successful. The paper calls for AI development and deployment frameworks that treat local validation, participatory design, continuous monitoring, and institutional accountability as minimum requirements for responsible global AI deployment.