Graph Neural Networks (GNNs) have emerged as an effective paradigm for learning from graph-structured data in domains such as social network analysis, bioinformatics, and recommendation systems. However, the performance of a GNN has remained highly sensitive to the selection of hyperparameters, including learning rate, hidden dimensions, aggregation functions, and regularization coefficients. Manual tuning and grid-based search strategies have often resulted in high computational cost and suboptimal configurations, which has limited scalability and reproducibility. The hyperparameter optimization problem in GNNs has posed a complex, non-convex, and high-dimensional search space. Conventional optimization approaches have struggled to adaptively explore this space, especially under limited computational budgets. As a result, GNN models have frequently suffered from overfitting, unstable convergence, or degraded generalization performance across different graph datasets. This study has proposed a bio-inspired metaheuristic optimization framework that has integrated population-based search principles with GNN hyperparameter tuning. A nature-inspired algorithm that has mimicked collective intelligence and adaptive behavior has guided the exploration and exploitation of the hyperparameter space. The proposed framework has encoded critical GNN hyperparameters as candidate solutions, which have been iteratively evolved using fitness feedback derived from validation accuracy and loss stability. The optimization process has been coupled with a training pipeline that has ensured fair comparison across candidate configurations. Experimental evaluation is conducted on benchmark graph datasets, including Cora, Citeseer, and Pubmed. The proposed method achieves peak classification accuracy of 88.0%, precision of 86.8%, recall of 86.5%, and F1-score of 87.0%, consistently outperforming Random Search, Bayesian Optimization, and PSO by 2–4.5%. Training time is reduced by approximately 10–15%, demonstrating both efficiency and scalability. Statistical analysis confirms that the improvements are significant, indicating robust generalization across datasets and stable convergence during hyperparameter optimization.
Agricultural productivity and public health are closely influenced by climatic variability, which has significantly affected both crop yield patterns and disease outbreaks. Extreme weather conditions have created the environmental factors that have increased the spread of vector-borne and climate-sensitive diseases. At the same time, the agricultural sector has depended on the crop yield forecasting systems that have supported food security planning. However, the existing analytical systems have treated agricultural prediction and health surveillance as two isolated domains. The absence of an integrated analytical framework has limited the ability of governments and agricultural agencies to anticipate the combined impacts of weather conditions on food production and public health risks. Conventional forecasting systems have analyzed yield patterns without considering the health indicators that have emerged from climatic fluctuations. To address this issue, this study has proposed a Weather–Agriculture–Health Integrated Mining Model (WAHIMM) that has combined crop yield prediction with climate-driven disease surveillance. The model has utilized weather attributes, agricultural yield records, and epidemiological indicators that have been collected from multi-source datasets. A hybrid learning pipeline has applied Random Forest regression for yield forecasting and Bayesian pattern mining for disease risk correlation, which has enabled the discovery of weather-dependent associations. The experimental evaluation demonstrates that the proposed Weather-Agriculture-Health Intelligent Mining Model (WAHIMM) achieves 97% accuracy, 95% precision, 95% recall, and 96% F1-Score while the mean absolute error decreases to 0.16. The framework significantly improves prediction performance when compared with the Regression Yield Model, the ML Crop Prediction model, and the Climate-Health Monitoring model. The integrated mining framework which analyzes the environmental variables that influence agricultural productivity and disease emergence provides reliable forecasting and early health risk detection.
Fuzzy logic deals with uncertainty, scalability, data integration, and inaccuracy that offers an appealing solution to Intelligent Transportation Systems (ITS), especially in traffic management in urban cities. This paper conducts a comparative study of five different fuzzy logic techniques, like Mamdani, Sugeno, Type-2, Adaptive Neuro-Fuzzy Inference System (ANFIS), and Genetic Fuzzy Systems (GFS), and evaluates their performance in a SUMO-MATLAB simulation framework. The results demonstrate that GFS has the shortest average wait time (29.90 seconds) and computational delay (0.08 milliseconds). Type 2 Fuzzy Systems, on the other hand, are better at dealing with sensor noise. Research has determined that a concentration on hybrid fuzzy approaches improves urban transportation.
Focal Cortical Dysplasia (FCD) is a common disorder causing drug- resistant seizures and it is usually managed surgically. The precise segmentation of lesions from MRI and other neuroimaging data is important to diagnose and plan surgery accurately. Typical U-Net models cannot capture the significant textures, irregular shapes, and delicate boundaries of FCD lesions. To alleviate these limitations, we introduce an advanced deep-learning model called ImReN2UNET that addresses FCD universal lesion segmentation using a GLCM-based loss function and improved residual block architecture. The introduction of GLCM-based loss function sharpens model localization and delineation by emphasizing textural features and spatial interdependence in lesions. This is a significant improvement in this field because identifying minute and complex lesions is of the utmost importance in medical imaging. The ImReN2UNET architecture runs on the robust presence of the nn-U-Net architecture and combines residual learning with texture information to segment the FCD lesion regions. The proposed method yields considerably more precise and reliable lesion segmentation by experimental comparisons against state-of-the-art segmentation techniques. This technique provides a powerful instrument for the diagnosis and assessment of FCD and thus informs clinical decisions for better outcomes for the patients.
Financial fraud in online credit card transactions poses significant challenges due to its increasing prevalence and the highly imbalanced nature of transactional data. This paper proposes a hybrid deep learning framework combining one-dimensional convolutional neural networks (CNN) and gated recurrent units (GRU) to effectively capture both spatial and temporal features of transaction sequences. Bayesian optimization is employed to fine-tune the model’s hyperparameters, improving detection performance without relying on synthetic oversampling. Evaluated on the widely-used European credit card fraud dataset, the proposed CNN-GRU model achieves superior results with an accuracy of 0.9996, an AUC-ROC of 0.9693, and an AUC-PR of 0.8709. These findings highlight the model’s robustness in identifying rare fraudulent transactions, outperforming several state-of-the-art methods and demonstrating the practical utility of deep learning combined with Bayesian optimization in fraud detection.
Sensitive data exposure on the Internet has increased due to the quick development of Internet-based services and applications. Users can save money and save time by using cloud storage instead of keeping local storage, however there are significant privacy and security issues. Data owners may create fine-grained access controls with Ciphertext-Policy Attribute-Based Encryption (CP-ABE), which guarantees that only authorized users can decode the data. This makes it a potential solution. For further protection, especially in cloud contexts, CP-ABE can be used with RSA to add temporal data sensitivity, fine-grained access control, and robust cryptographic foundations. In addition to enforcing stringent access control, this study suggests a unique CP-ABE with RSA (CARSA) approach that identifies the users participating in the decryption process. The proposed CARSA system is shown to be more efficient by experimental assessment, with encryption and decryption timings of 1610 ms and 1098 ms for a 256-bit key length, respectively. By contrast, current methods take 1134 ms and 1691 ms, respectively. According to the findings, CARSA offers improved speed in cloud-based settings while guaranteeing strong data security
In this study, two machine learning models, Long Short Term Memory (LSTM) and BERT are used to predict intensifiers in Malayalam sentences. Both models were trained to detect intensifiers using part-of-speech (POS) tags, and BERT regularly outperformed more straightforward models like Naive Bayes (NB) and Support Vector Machines (SVM) in terms of metrics like accuracy, precision, recall, and F1 score. In contrast to LSTM, which was effective but suffered from overfitting as demonstrated by the comparison of training and validation losses, BERT’s self-attention mechanism allows it to grasp intricate associations between words. LIME and SHAP visualisations further clarified the role that individual words played in sentiment classification. The results demonstrate BERT’s better performance in handling the complex intensifier prediction problem. With an emphasis on its attention process as examined by BERTology, this study demonstrates BERT’s proficiency in predicting intensifiers in Malayalam sentences. Compared to models like LSTM, BERT is far better at capturing intricate interactions between words, such intensifiers and their surrounding context, thanks to its multi-layered design and self-attention mechanism. With early layers focussing on local linkages and subsequent layers collecting broader, more global dependencies, the attention heads in BERT enable the model to concentrate on certain tokens inside the phrase. Because of its capacity to focus on various phrase components, BERT is able to comprehend the nuanced relationships between intensifiers and adjectives, which results in extremely accurate predictions at the sentence and token levels.We can observe how BERT gradually improves its comprehension of the input by visualising the attention weights across layers. This allows it to create rich contextual representations, which are essential for tasks such as sentiment analysis. This knowledge of BERT’s attention mechanism explains why it performs better than other models in recognising intensifiers and determining sentiment intensity.
Identification of Feature interaction is the challenging task particularly for high dimensional database in the research field of feature selection which plays a major role in Machine Learning. “Effective RNN based Feature Interaction Model using Cloud Databases” is a Recurrent Neural Network based approach in the Cloud environment which detects positive feature interaction. The proposed RNN based Feature Interaction model consists of Feature Interaction Identifier and Redundant Feature Interaction Remover. The Feature Interaction Identifiers identifies the Feature Interaction pair in the given dataset which is retrieved from Cloud database. The Redundant Feature Interaction Remover removes the duplicate Feature Interaction Pairs which are produced from the Feature Interaction Identifiers and achieves Positive Feature Interaction. The four Medical Databases obtained from AWS Cloud is used to evaluate the performance of the proposed Feature Interaction Model. The experimental and evaluation results obtained on these Cloud databases show that a proposed RNN based Feature Interaction Model performs feature interaction with significant improvement in classification accuracy better than conventional Feature Selection Models.
Brain tumor segmentation has remained a critical task in medical image analysis, as accurate delineation has directly supported diagnosis, treatment planning, and clinical decision-making. Conventional deep learning approaches have achieved notable success; however, they have often struggled with limited robustness when facing heterogeneous tumor shapes, intensity variations, and imaging noise across multimodal MRI data. Existing segmentation frameworks have relied heavily on deterministic optimization strategies that have suffered from premature convergence and reduced generalization. These limitations have affected segmentation consistency, particularly in complex tumor boundaries and low-contrast regions, which have demanded adaptive and globally optimized solutions. This study has presented a quantum-inspired evolutionary framework that has integrated probabilistic quantum representation with an evolutionary optimization mechanism. The proposed framework has encoded segmentation candidates using quantum bits that have allowed superposition-based exploration of the solution space. An evolutionary update strategy has guided probability amplitudes toward optimal states, while a convolutional segmentation backbone has extracted hierarchical spatial features. A fitness-driven selection process has refined candidate solutions that have maximized region similarity and boundary accuracy. The training process has incorporated adaptive mutation and crossover operators that have preserved diversity and stability during convergence. Experimental evaluation demonstrates that the proposed method achieves superior performance on benchmark brain MRI datasets. The Dice similarity coefficient reaches 0.91–0.93 across modalities, while the Jaccard index ranges from 0.81– 0.84. Sensitivity achieves up to 0.92, and specificity remains high at 0.90–0.94. Overall accuracy ranges between 0.89–0.94, surpassing conventional CNN-based and evolutionary baselines. Visual inspection confirms precise tumor boundary delineation, particularly in infiltrative regions, indicating enhanced robustness under noise and intensity variations.
Medical image fusion has played a critical role in clinical diagnosis by integrating complementary information from multi modal sources such as MRI, CT, and PET. Conventional fusion techniques have suffered from information loss, spectral distortion, and weak adaptability under complex anatomical variations. Recently, deep learning and fuzzy inference approaches have emerged as promising solutions, yet they have remained sensitive to parameter initialization and local optima. Existing deep neuro-fuzzy fusion models have exhibited limited robustness due to static membership functions and suboptimal rule optimization. These limitations have resulted in blurred edges, reduced contrast preservation, and unstable fusion quality across heterogeneous imaging modalities. The lack of adaptive optimization has restricted their generalization in real clinical environments. This work has proposed a swarm-enhanced deep neuro-fuzzy system for multi modal medical image fusion. A deep neuro-fuzzy architecture that has integrated convolutional feature extraction with fuzzy inference has been developed. Swarm intelligence that has included particle-based optimization has been employed to adaptively tune fuzzy membership parameters and rule weights. Feature learning that which has captured spatial and textural cues has been followed by a fuzzy decision layer that which has modeled uncertainty and nonlinearity. The fusion strategy has combined salient features using optimized fuzzy rules, while reconstruction that which has preserved anatomical consistency has been performed. Experimental evaluations are conducted on standard multi modal medical image datasets. The proposed system achieves higher entropy (up to 7.11), structural similarity index (up to 0.94), edge preservation index (up to 0.88), peak signal-to-noise ratio (up to 33.8?dB), and mutual information (up to 2.68) compared with conventional deep learning and fuzzy-based fusion methods. Visual analysis demonstrates that clinically relevant structures are better preserved while noise and artifacts are significantly reduced. The swarm optimization that which guides parameter learning improves convergence stability and fusion consistency across modalities.
India's software sector contributes significantly to GDP and has a substantial share of the export market. This study ranks the performance of selected Indian software companies according to their technical efficiency. The company’s efficiency is evaluated using the input-oriented DEA technique over an extended period, from 1992 to 2021. Performance trends are generated. Companies that perform best are compared with those that perform poorly or inefficiently. To increase efficiency, firms can follow the business strategies, techniques, ethics, and best practices of efficient companies. The results of the present study could assist in managerial decision-making.
In the field of medical data analysis, the major challenges are high dimensionality and complexity due to temporal behavior of medical datasets. Feature selection is crucial to overcoming these challenges since it enhances interpretability, reduces processing expenses, and boosts model performance. To analyze the medical records, it is very essential to determine the most potential features that contribute more in classification or diagnosis of disease especially in the medical in the Medical Information Mart for Intensive Care III (MIMIC-III) dataset. This paper, highlights the importance of robust feature selection by developing a novel Hybrid feature selection framework that combines ReliefF and Ant Colony Optimization with Graph Neural Networks (ACO-GNN). The first step in the suggested approach is Ant Colony Optimization (ACO), which creates candidate feature subsets by effectively exploring the combinatorial feature space by mimicking pheromone-guided search behavior. Then, using graph-based representations of clinical variables, such as correlations between lab tests, drugs, and vital signs, Graph Neural Networks (GNNs) are used to model intricate, non-linear interactions among medical aspects. In order to ensure robustness and interpretability, ReliefF is used to rank and improve features by assessing their capacity to distinguish between patient outcome classes. The hybrid approach significantly outperforms conventional feature selection techniques like K-Nearest Neighbors with ReliefF (KNN-ReliefF) and XGBoost with SHAP Feature Importance (XGB-SHAP) in through tests on the MIMIC-III dataset, predictive performance indicators such as precision, accuracy, F1-score, recall, and AUC-ROC. The selected feature subsets offer clinically meaningful insights into critical factors influencing patient outcomes in intensive care, underscoring the potential of the ACO-GNN-ReliefF method for advancing predictive analytics and clinical decision support systems in healthcare.
The rapid adoption of the data-driven healthcare analytics has raised serious concerns regarding the patient privacy, data integrity, and collaborative intelligence across distributed medical institutions. Traditional centralized learning approaches have relied on extensive data sharing that has increased the risk of data leakage and regulatory noncompliance. Federated learning has emerged as a promising paradigm that has enabled collaborative model training without direct data exchange. However, the presence of unreliable or malicious participants has limited its practical deployment in real-world healthcare environments. Although federated learning has preserved data locality, it has not fully addressed the issue of trust among participating clients. The contribution of low-quality or adversarial updates has degraded the global model performance and has compromised the clinical reliability. Existing aggregation strategies have ignored behavioral uncertainty and contextual trust, which has resulted in biased or unstable healthcare predictions. This study has proposed a trust-aware federated learning framework that has integrated soft computing techniques for adaptive client evaluation. A fuzzy logic-based trust model has assessed each participant using the historical update consistency, model divergence, and communication reliability. The trust scores that have been computed have dynamically weighted the local updates during aggregation. A privacy-preserving mechanism that has incorporated differential noise has further strengthened data confidentiality. The framework has been validated using distributed healthcare datasets that have represented diagnostic classification tasks under heterogeneous data distributions. The experimental evaluation demonstrates that the proposed trust-aware federated learning framework achieves a classification accuracy of 0.94 and an F1-score of 0.94 at 200 iterations, which outperforms Federated Averaging, Differentially Private Federated Learning, and Trimmed Mean aggregation by margins of 10–15%. The framework reduces convergence time to 95 rounds, compared with 140–175 rounds for existing methods. These results confirm that trust-guided aggregation improves robustness, accelerates convergence, and preserves privacy in distributed healthcare analytics.
Wearable physiological sensors have enabled the continuous acquisition of cardiac signals that has supported early health monitoring outside clinical environments. However, the variability, noise, and temporal complexity of wearable signals have limited the reliability of conventional analytical models. Existing approaches have struggled with capturing both long-term temporal dependencies and localized morphological patterns within the same framework, which has reduced their clinical applicability for early cardiac anomaly detection. The accurate identification of early-stage cardiac anomalies from wearable signals has remained challenging due to signal artifacts, inter-subject variability, and the imbalance between normal and abnormal patterns. Traditional machine learning models have relied on handcrafted features that have failed to generalize across diverse populations. Deep models without interpretability have also raised concerns regarding trust and deployment in real-world monitoring systems. This study has proposed a hybrid attention-guided LSTM–CNN architecture that has integrated temporal sequence learning with spatial feature extraction. A convolutional neural network has extracted localized signal characteristics, while a long short-term memory network has modeled sequential dependencies that have evolved over time. An attention mechanism that has selectively emphasized clinically relevant segments has improved feature weighting and interpretability. The model has trained on preprocessed wearable cardiac signals that have undergone normalization, denoising, and segmentation. Experimental evaluation has demonstrated that the proposed model has achieved superior detection accuracy, sensitivity, and specificity compared with baseline CNN and LSTM models. The attention module has contributed to improved robustness under noisy conditions and has enhanced early anomaly recognition. Statistical analysis has confirmed consistent performance gains across multiple evaluation folds, indicating reliable generalization.
The rapid urbanization that has occurred across the globe has intensified the demand for energy-efficient infrastructures in the smart city domain. Conventional control strategies have struggled to address the dynamic, nonlinear, and uncertain nature of urban energy systems. Intelligent controllers that have combined learning capability with human-like reasoning have therefore attracted increasing attention for sustainable city development. Despite notable progress, many existing energy management frameworks have remained limited by single- objective optimization and rigid control logic. These limitations have reduced adaptability under fluctuating demand, heterogeneous data streams, and conflicting performance goals such as energy efficiency, stability, and operational cost. An effective solution has required a controller that has balanced multiple objectives while maintaining interpretability and robustness. This study has proposed a multi- objective neuro-fuzzy controller that has integrated fuzzy inference with neural learning for smart city energy management. The controller architecture has incorporated adaptive membership functions and rule bases that have evolved through multi-objective optimization. Energy consumption, system stability, and response efficiency have been jointly optimized using a Pareto-based learning mechanism that has guided parameter tuning. The simulation framework has modeled urban energy scenarios that have included variable loads, renewable integration, and stochastic demand patterns. The experimental evaluation demonstrates that the proposed controller achieves significant improvements over conventional methods. Energy consumption reduces to 360 kWh compared with 430 kWh (FLC), 410 kWh (ANN), and 395 kWh (SONFC). Stability index increases to 0.93, and response time decreases to 1.2 s. Energy savings reach 21.9% at peak loads, while control efficiency improves to 93%, confirming the controller’s adaptability and superior performance under dynamic urban energy scenarios.
Agricultural intensification has increasingly relied on the widespread application of agrochemicals, which have often dispersed beyond targeted farmlands through atmospheric drift. This environmental exposure has raised concerns regarding respiratory health among rural populations. Previous environmental surveillance systems have rarely integrated meteorological variables with clinical records that describe respiratory disease patterns. Therefore, the lack of an integrated analytical model has limited the ability to understand the association between agrochemical drift and rural respiratory illness. This study has proposed a Multivariate Agrochemical Drift Impact Mining Model (MADIMM) that which integrated weather attributes and rural clinical records to estimate the relationship between agrochemical dispersion and respiratory disease occurrence. The framework has utilized multivariate data mining techniques that which analyzed temperature, humidity, wind speed, precipitation, and seasonal spraying patterns together with hospital respiratory admission data. The preprocessing stage has included normalization, missing value imputation, and feature correlation filtering. Subsequently, the MADIMM classifier has applied ensemble learning that which combined Gradient Boosting and Random Forest models to extract environmental–clinical correlations. The experimental evaluation is showing that the proposed MADIMM framework achieves 96% accuracy, 95% precision, 94% recall, 95% F1 score, and 97% AUC in respiratory disease prediction. The model improves classification accuracy by approximately 10–13% compared with Random Forest, Support Vector Machine, and Gradient Boosting models. The environmental drift exposure modeling captures atmospheric dispersion patterns that which significantly improve the prediction of respiratory disease risk in rural agricultural regions.
Medical image analysis is very essential for health care sectors with early identification of illness, advanced features as very effective diagnostics system. Because, medical image modalities of mammogram images to screening the breast experiment of radiologist taken procedures to given data for diagnostic with low radiation of X-ray images. Machine learning methods for support vector machine utilizing image enhancement of detect breast cancer. MIAS data applied by the 322 images are analysis to image enhancement of eliminating noise with regard to filtering methods from feature extraction. The objective methods as following from best quality to identify with best models’ findings that the ML techniques. Towards zernike moments and mahotas were used for obtaining severely features to mammogram images, when that data’s such that benign or malignant to SVM approaches with linear over radial basis functions to that kernel. While optimization of feature selection and classification, as conquest to levied for peak signal to noise ratio, signal to noise ratio and mean square errored to assistance the potency of methods. Too feature extraction of zernike moments with SVM classification provides that very high performance in identifying indicators of cancer of breasts. Proposed methods as machine learning techniques like that SVM based approaches these methods for finding the best methods to performance metrics with quality of image enhancement for PSNR, SNR and MSE also overall accuracy with better models from zernike moments of feature extraction. Finally, get the results to target that better outcome for image enhancement of breast cancer identification.
The performance of modern websites has become a critical factor in determining user engagement, service reliability, and computational efficiency. Background studies have shown that increasing traffic load, dynamic content delivery, and heterogeneous user behavior have significantly reduced responsiveness and scalability of web systems. The problem of inefficient resource utilization and delayed response time has remained persistent, especially in cloud-hosted and data-intensive web environments. To address this issue, a Hybrid Deep Reinforcement Learning and Gradient Boosted Optimization (HDRL-GBO) method has been proposed to improve website performance through adaptive resource allocation and predictive load balancing. The proposed HDRL-GBO method integrated deep reinforcement learning for real-time decision optimization and gradient boosted machine learning for performance prediction. The reinforcement component dynamically adjusted caching strategies, request routing, and server allocation, while the boosting model estimated workload patterns and latency variations. The system has been trained using historical traffic logs and real-time simulation data. The proposed HDRL-GBO framework has significantly reduced average latency to 85–140 ms, compared to 105–240 ms in baseline methods. Throughput has improved to 520–600 req/s, while existing methods achieve 340–470 req/s. CPU utilization has reached 92% balanced efficiency, outperforming EHGA at 90% and PSO at 88%. These results confirm that the proposed system has improved performance stability, scalability, and response efficiency under dynamic web traffic conditions.
Medical image analysis has improved significantly using deep learning, but accurate segmentation remains challenging due to variability in disease patterns, anatomical structures, and image quality. This paper proposes a Hybrid CNN-U-Net Model with Adaptive Attention Mechanism (AAM) for automated brain tumor segmentation. The methodology combines CNN-based hierarchical feature extraction with U-Net’s segmentation capability, enhanced by AAM that dynamically concentrates on salient tumor regions. Evaluated on BraTS 2021 dataset, this proposed model here achieves a Dice coefficient of 93.1±1.4 with 95% confidence interval [91.5, 94.7], representing a 1.1-point improvement over the Li et al. baseline (92.0±1.5). Ablation studies isolate the contribution of each component. Statistical hypothesis testing confirms significant improvements over standard U-Net and U-Net + Attention baselines (p < 0.001). The model demonstrates greater performance in segmenting Whole Tumor (WT), Tumor Core (TC), and Enhancing Tumor (ET) regions. Per-class Dice scores are: WT = 92.8±1.2, TC = 89.9±1.8, ET = 85.6±2.3. Cross-dataset evaluation on BraTS 2020 shows generalization capability (Dice = 90.1±2.1, degradation = -2.9%).
This paper introduces an attention-guided joint CNN–neuro-fuzzy learning framework for offline handwritten signature verification, in which convolutional feature learning and fuzzy membership optimization are performed simultaneously rather than through a conventional sequential pipeline. Grad-CAM–based attention cues are incorporated to guide the adaptive update of fuzzy membership functions, allowing the model to emphasize discriminative signature regions during decision making. To improve robustness across scripts and writing styles, a domain-adaptive training strategy is adopted, enabling effective generalization across English, Bengali, and Hindi signatures. Experiments conducted on the CEDAR, GPDS, and BHSig datasets demonstrate consistent improvements in cross-dataset transfer accuracy over baseline CNN, fuzzy-only, and conventional CNN–fuzzy approaches. The proposed joint learning strategy enhances model interpretability by linking attention-derived feature importance with fuzzy inference rules, while domain adaptation contributes to improved resilience against variations in script, writer style, and acquisition conditions. Overall, the framework provides a step toward more explainable and multilingual signature verification systems.
