The exponential growth of high-dimensional data across domains such as healthcare, finance, and industrial automation necessitates robust models capable of both accurate predictions and explainable decision-making. Traditional machine learning models, such as Support Vector Machines (SVM), Random Forests, and ensemble techniques, often struggle to generalize effectively when feature spaces become large and noisy. Deep learning models such as CNNs, LSTMs, and Transformers, while powerful, typically operate as black boxes, providing limited interpretability and weak uncertainty quantification. To address these challenges, this study introduces the MetaXAI Adaptive Learning Model, a meta-learning-based explainable artificial intelligence framework designed for complex algorithmic decision-making in high-dimensional environments. MetaXAI integrates autoencoder-based dimensionality reduction, attention-driven embedding, and a Model-Agnostic Meta-Learning (MAML) engine with Gradient Episodic Memory (GEM) for task adaptation. Furthermore, it combines SHAP, LIME, and counterfactual explanation modules to enhance interpretability. Results demonstrate that MetaXAI achieves an accuracy of 98.91%, outperforming traditional models, such as SVM (91.3%), XGBoost (93.8%), and Transformer (94.8%). It also exhibits superior explainability (fidelity 98.2%) and uncertainty calibration (Expected Calibration Error 0.9%). Federated deployment further ensures privacy preservation with minimal communication overhead. The findings suggest that MetaXAI bridges the gap between predictive performance and interpretability, making it suitable for critical domains where both accuracy and transparency are paramount.
The evolution of 6G networks brings unprecedented connectivity and processing capabilities to Internet of Things environments, yet it also raises more challenges in data security, intrusion detection, and computational efficiency. In this paper, a complete end-to-end framework with the inclusion of Artificial Intelligence, blockchain technology, and novel encryption techniques is proposed to address these challenges in 6G-IoT networks. Then, the AI-Powered Cybersecurity Events Dataset is utilized first, wherein network traffic is normalized through Min-Max normalization to normalize heterogeneous features to a uniform scale. A Gated Recurrent Unit-based neural network is then trained on this normalized data to detect real-time intrusions by learning complex temporal dependencies. Upon detecting anomalies, a blockchain layer is called to execute smart contracts that apply automatic security measures, e.g., quarantining affected nodes. For efficient processing, the architecture accommodates GRU-assisted task offloading optimization on a multi-factor delay, energy, and network load optimization model. Blockchain smart contracts manage load balancing and delegation verification autonomously without any central authority. Homomorphic encryption and proxy re-encryption also ensure data confidentiality in edge computing and secure multi-party computation. All encryption and task offloading operations are permanently stored onto the blockchain, ensuring system-wide transparency and auditability. Federated learning support is built-in to enable privacy-preserving decentralized AI training from distributed edge nodes. The entire framework is implemented in Python using TensorFlow/Keras for the GRU model and web3.py for blockchain-related interactions. Experimental measurements demonstrated the efficiency and stability of the system by exhibiting a percentage of accuracy of 98.9% in intrusion detection, justifying the applicability for future 6G-IoT secure and scalable operations. The experiments utilize the AI-Powered Cybersecurity Events Dataset containing approximately 1.5 million network flow records with class distribution of 52% benign traffic, 21% DoS, 15% probing, and 12% privilege escalation events.
Mental health sentiment expressed via bilingual and code-mixed text, such as Hindi text written in Latin script, presents some unique challenges to computational understanding of language. This paper proposes GAEn/MGT (Generative Auto-encoder Network / Multi-generative Transformer), an integrated generative-transformer model used for robust sentiment analysis task for mental health discourse, which has bilingual and transliterated data. In proposed model, the generative reconstruction, variational regularization and supervised classification are all embedded into a holistic hybrid learning objective, which enables high-quality cross-linguistic alignment and semantic preservation. Generally, GAEn/MGT uses the multilingual transformer backbones (mBERT, XLM-R, MuRIL, and IndicBERT) with transliteration-consistent embeddings and domain alignment mechanisms in order to deal with noisy, code-mixed and emotionally complex text. Experiments on large scale bilingual mental health data sets (Twitter, Reddit, ISEAR, Sentiment140, and Hinglish corpus) have shown that GAEn/MGT achieves 0.920 accuracy and 0.908 macro-F1 (Precision = 0.910, Recall = 0.905). Stratified bootstrap confidence intervals are used to verify the robustness of these results (macro-F1 95% CI: [0.897, 0.916]), and this improvement over the best performing baseline XLM-R is statistically significant (p ≤ 0.01). Beyond achieving quantitatively higher results, the model is also more interpretable thanks to reconstruction and latent visualization, which reveals greater insight into variance in emotion across languages. The results signal GAEn/MGT as an efficient and inclusive system, which is ethically responsible about early detection, emotional monitoring and digital-psychiatric research in multilingual societies.
Face recognition systems often fail to distinguish monozygotic twins, commonly called identical twins, due to their high facial similarity. Specifically, changes in illumination or expression reduce the discriminative ability of such systems further. Hence, this research introduces a novel twin face recognition methodology that combines Super-Resolution Convolutional Neural Networks (SRCNN) and Multi-Orientation-Summed Gabor Local Binary Patterns (MSGLBPs) to achieve twin recognition across four test cases, which involve real-time constraints, such as illumination variations and changing facial expressions. At first, the SRCNN module of the proposed method enhances the high-frequency details and the texture quality in the face images of twins, so as to improve recognition under such constraints. Subsequently, the subtle biometric differences between the twin faces are extracted by capturing local texture variations across multiple Gabor orientations using MSGLBP. Finally, identity recognition is done using Support Vector Machine (SVM). Experimental outcomes confirm that a recognition accuracy and an [Formula: see text]1-score of up to 98.48% and 0.98, respectively, are achievable with this method using the ND-TWINS-2009-2010 dataset. This research contributes to advancing face recognition technology and ensuring security in Human Computer Interaction (HCI), particularly in challenging scenarios involving identical twins.
The necessity to identify cyberattacks in the systems of the Industrial Internet of Things (IIoT) has become topical since the latter is a key element of the Industry 4.0 services, such as smart manufacturing, energy control, and industrial communications. The rapid growth in the number of connected devices and cyberphysical objects has exposed IIoT environments to sophisticated cyberattacks, including DDoS attacks, malware, and zero-day attacks, potentially leading to disruption of operations, data integrity and data safety. The existing intrusion detection systems have severe problems with high complexity of calculation, low ability to adapt to novel attacks, and low detection. To overcome these issues, this paper will propose a hybrid cybersecurity solution that will use the Dandelion Optimization Algorithm (DOA) to select the features, in addition to a new Longformer Vision Randomized Network (LoV-RaNet) to classify the attack. The selected features are most relevant and discriminative as DOA reduces redundancy, hence lowering the computational cost, whereas LoV-RaNet utilizes long-range attention and randomized residual connections to accurately identify known and unknown attacks. The proposed framework enables the scalable, efficient and real-time IIoT intrusion detection. The experimental results also show that LoV-RaNet is more effective, with the accuracy of 0.990, the precision of 0.985, the recall of 0.982, and [Formula: see text]1-score of 0.988 being higher than the conventional intrusion identification models.
A novel hierarchical evolutionary feature selection (HEFS) framework is proposed to enhance interpretability and predictive performance in academic performance analysis. The framework addresses the challenge of selecting meaningful features from high-dimensional lifestyle data while preserving psychological validity. Initially, features are organized into semantically coherent clusters through hierarchical agglomerative clustering, employing correlation-based distance metrics and Ward’s linkage to ensure logical groupings. These clusters are subsequently optimized using an evolutionary algorithm (EA) designed to balance predictive accuracy with interpretability, guided by established behavioral theories. The selected feature subsets are further validated against recognized psychological constructs to ensure theoretical relevance, thereby integrating data-driven insights with domain knowledge. The refined feature set is then incorporated into interpretable machine learning models to support transparent and accountable decision-making. Experimental results indicate that HEFS outperforms conventional feature selection methods in terms of both predictive performance and explainability. This study contributes a principled and scalable approach to feature selection that reconciles statistical rigor with psychological grounding, offering practical value for educational analytics and related domains. While the primary validation uses a single-institution dataset, cross-program generalization experiments (AUC-ROC = 0.841 ± 0.004) indicate robustness across diverse curricular structures, though multi-institutional replication remains necessary.
The flexible job shop scheduling problem with automated guided vehicles (FJSP–AGV) couples production and transport decisions, making scheduling and energy management computationally challenging. Conventional genetic algorithms apply a single decoder throughout the search and thus cannot adapt when instance characteristics or battery constraints change. We propose a portfolio–island decoding genetic algorithm (PID-NSGA-II) that shifts the focus from modifying evolutionary operators to learning which decoding strategies work best. Five heterogeneous decoders run in parallel on separate islands, and an upper-confidence-bound multi-armed bandit measures each island’s contribution to makespan improvement and adaptively reallocates population resources, automatically balancing exploration and exploitation. The framework is tested under two settings—pure makespan minimization and energy-aware scheduling with AGV battery considerations. Experiments on benchmark datasets show that PID-NSGA-II consistently improves solution quality and stability compared with single-decoder genetic algorithms, with greater gains when energy constraints are present. Adaptive learning of decoders delivers more robust scheduling decisions for complex FJSP-AGV environments and provides a scalable platform for smart manufacturing applications, achieving up to 25 % makespan reduction and substantial improvements in AGV battery levels across small, medium and large problem instances.
A set of jobs has to be scheduled on a single machine starting from a non-zero time. The processing time of one job from the set is a quadratic non-decreasing function of its start time, while the processing times of the other jobs are proportional non-decreasing functions of their start times. The objective is to minimize the makespan. We present a criterion for optimally scheduling a job with proportional processing time alongside the job with quadratic processing time. We identify some polynomially solvable cases and provide for the remaining cases an upper bound on the starting time of the job with quadratic processing time. Additionally, we establish the NP-hardness of this problem and propose a pseudo-polynomial-time optimal algorithm and a fully polynomial-time approximation scheme.
The growing impact of artificial intelligence on everyday life has fueled the emergence of Human-Centered Artificial Intelligence (HCAI), an approach aimed at developing intelligent systems that not only emulate and extend human capabilities but also foster meaningful collaboration, societal integration, and ethical responsibility. In this work, we present a comprehensive survey of the role of argumentation in HCAI, examining its theoretical foundations, methodologies, and applications, as well as its potential to enhance communication, informed decision-making, and responsible human-AI interaction across diverse domains. Building on this analysis, we introduce the Argumentation-based Proof-Event Calculus (APEC), a formal and aligned framework designed to support trustworthy collaboration between humans and AI agents in goal-oriented decision processes. The proposed APEC framework advances HCAI by explicitly incorporating temporal dynamics, evaluating both individual and collective agent contributions in multi-agent systems, and supporting the iterative refinement and evolution of decisions toward well-justified outcomes.
Mislabeling is a challenging problem in self-labeled methods utilizing labeled and unlabeled samples to train classifiers. For overcoming misprediction, the existing self-labeled methods employ ensemble classifiers to improve predictive performance on unlabeled samples or employ heuristic strategies to find correctly predicted samples. Yet, they experience the following issues: (a) adopted ensemble classifiers employ random resampling or entire labeled samples without considering training sample quality and usefulness; (b) adopted heuristic strategies for selecting high-confidence pseudo-labeled samples explicitly rely on assumptions about geometric and class relationships; and (c) a few adopt ensemble classifiers and heuristic strategies to overcome misprediction. Inspired by the state-of-the-art sample subspace optimization with accelerated binary particle swarm optimization (SSO-ABPSO), an ensemble self-labeled method (ESLM-SSO) based on sample subspace optimization with accelerated binary particle swarm optimization is proposed against the above issues. The novelties of ESLM-SSO are as follows: (a) a SSO-ABPSO-based ensemble classifier considering training sample quality and usefulness is proposed to predict unlabeled samples, and (b) a SSO-ABPSO-based heuristic strategy that evaluates pseudo-labeled samples based on empirical subset performance, rather than explicit geometric assumptions, is proposed to identify correctly predicted samples. Experiments demonstrated that ESLM-SSO outperforms seven state-of-the-art self-labeled methods in improving two classifiers on extensive datasets.
Predicting student academic performance has become increasingly vital in the field of educational data mining, as institutions seek data-driven strategies to enhance learning outcomes. However, many existing models rely solely on behavioral indicators or static features, often overlooking the role of time and context in shaping learning behavior. This limitation reduces predictive accuracy and adaptability in academic environments. To address this challenge, this study introduces EduFuseNet, a hybrid deep learning framework that integrates behavioral and spatiotemp oral data for accurate classification of student performance. The workflow begins with data collection from a Student Academic Performance dataset, comprising both behavioral metrics and spatiotemp oral information. The raw data undergoes preprocessing, including missing value imputation, one-hot encoding of categorical variables, and min-max scaling of numerical features. The processed data is then passed through two specialized branches: a Tabular Neural Structure-Aware (TabNSA) module that captures complex interdependencies within behavioral data, and a Spatiotemp oral Transformer module that models temporal and sequential patterns in learning activities. The feature embeddings from both branches are fused and passed through fully connected layers to generate predictions across five academic performance bands, enabling precise classification and early risk identification. EduFuseNet achieved an accuracy of 99.00%, with a precision of 99.04%, recall of 99.00%, and F 1-score of 99.01%, reflecting strong and reliable predictive performance. By leveraging both behavioral and temporal learning indicators, the model serves as an effective tool for early academic monitoring and intervention.
This work developed a new predictive system using an innovative deep learning approach for traffic management using high-resolution aerial images. The proposed method leverages a novel 3D Transformer-based You Only Look Once v9 (3D-TYolov9) model for the detection and tracking of vehicles from converted 3D image frames (feature 1, f1). The developed 3D-TYolov9 is defined as a system that likely combines the established You Only Look Once v9 (YOLOv9) object detection framework with extensions for 3D data processing, specifically incorporating a Transformer-based module for enhanced feature analysis and tracking within 3D space. Subsequently, 2D features reflecting real-time traffic behavior, such as speed and position (f2), are extracted. To enhance accuracy further, diverse external 1D contextual features (f3), including weather conditions and event disruptions, are integrated. These multi-dimensional features (f1, f2, and f3) are then processed by a Multi-Dimensional Convolution with Sparse Attention-based Dilated Densenet (MDC-SADDNet) for robust traffic forecasting. This architecture effectively captures complex patterns, enabling significant improvements in future traffic management, smart routing, and urban mobility planning. Performance validation confirms the developed model’s effectiveness over existing methods.
Image quality enhancement is a critical task in computer vision and image processing, particularly in poor-lighting conditions where noise and distortion often degrade the perceptual quality. This paper presents a novel approach for image quality enhancement using a cosine quality evaluator custom network for indoor object detection. Adaptive filtering techniques, such as Cosine Adaptive Enhanced Block-Matching and 3D filtering, focus on denoising to effectively reduce noise while preserving key image features. The Blind/Reference Image Spatial Quality Evaluator score allows for effective evaluation without a reference image by examining statistical qualities and distortion characteristics to offer a no-reference measure for evaluating image quality. Deep custom convolutional neural networks are employed after denoising, classifying images based on their enhancement score to enhance the quality. This process ensures that important details, such as edges and textures, are retained while improving overall image clarity. Despite the difficulties caused by high computational costs, real-time processing is achieved through efficient denoising and quality assessment methods. The combination of these methods allows for effective image enhancement, generally in complex situations involving blur, noise, and low-light conditions, with improved processing efficiency and perceptual quality. At 99.53% accuracy and a 99.62% F1-score, the proposed Blind Quality Evaluator Custom Network (BQECN) approach achieves exceptional results. The model shows robustness with an accuracy of 99.86% and an F1-score of 99.76% in 10-fold cross-validation. These outcomes demonstrate the method's ability to handle blurred and noisy images.
The rapid growth of Industry 4.0 and the rise of smart cities have generated massive, heterogeneous datasets across transportation, healthcare, energy, and public services. Harnessing these data streams for efficient and sustainable urban management requires artificial intelligence (AI) solutions that are both accurate and transparent. Traditional AI models often function as opaque black boxes, limiting their acceptance in high-stakes urban governance. To address this challenge, we propose an Explainable Artificial Intelligence (XAI) framework that integrates scalable data analytics, interpretable machine learning, and visualization modules for real-time decision support in smart city environments. Built on distributed big data infrastructures, the system processes high-velocity data streams and employs feature attribution, rule-based explanations, and model-agnostic interpretability techniques to generate human-understandable insights. The framework is evaluated using the Smart Cities Index dataset from Kaggle, encompassing multi-domain indicators covering infrastructure, mobility, environmental performance, and governance metrics, with experiments conducted across classification and ranking tasks to assess both predictive performance and explanation fidelity. Case studies in smart energy management, intelligent traffic control, and public safety demonstrate that the framework enhances prediction accuracy while improving interpretability, stakeholder trust, and decision transparency. By bridging the gap between high-performance AI and human interpretability, the proposed approach supports accountable, fair, and sustainable smart city operations aligned with Industry 4.0 principles.
Natural language processing encounters major difficulties when processing low-resource languages because technical fields require specialized terminology and lack sufficient annotated data. The performance of transformer-based multilingual models depends on their ability to operate in different domains and their available resources for training. The researchers developed a transformer-based framework that uses XLM-RoBERTa (XLM-R) to process languages that have limited resources for their needs. The proposed methodology includes systematic data preprocessing methods that create low-resource data through data augmentation and implement a two-stage domain adaptation approach that begins with pretraining on unlabeled technical text before moving to task-specific classification fine-tuning. The FLORES-200 multilingual dataset provides a structured benchmark assessment that tests low-resource conditions through its multiple testing methods. Model performance assessment uses standard evaluation metrics, which include accuracy and precision and recall and F1-score, together with confusion matrix analysis and evaluation curve measurements. The experimental results show that the proposed approach achieved 0.98889 accuracy, 0.98925 precision, 0.98889 recall, and 0.98889 F1-score, which demonstrates strong accurate classification abilities. The study shows that multilingual transformer models, which use domain adaptation and data augmentation methods, deliver effective solutions to data scarcity challenges while improving performance in technical fields.
The increasing number of web attacks in a network environment poses substantial cyber risks to the network devices and systems, particularly in Industry 4.0, where interconnected smart devices and critical infrastructures are highly vulnerable. Conventional Intrusion Detection Systems (IDS) rely on predefined attack signatures and rule sets, which make them ineffective against multi-vector and social engineering attacks. This limitation can increase the likelihood of true negative incidents in the networks, potentially affecting the entire network. To address this limitation, a DL-IDS (Dual-Layer Intrusion Detection System) architecture is proposed, with effective threat detection with minimal computational overhead. The proposed architecture consists of two layers: Hybrid Threat Detection Model (HTDM) and Confirmatory Ensembled Models (CEM). HTDM monitors the traffic patterns to detect anomalies and flags suspicious packets as high Alert (hA). These hA packets are forwarded to the CEM layer, which uses specialized ensemble models for different attacks such as DDoS, Patator, and Web Attacks. This structured dual-layer approach enhances accuracy, reduces false positive and computational overhead. The proposed model is trained on the CICIDS-2017 dataset. HDTM achieved 99.92% accuracy, while CEM models of DDoS, Patator and Web Attacks achieved 98.57%, 98.97% and 98.91% respectively. The proposed architecture reduces the computational overhead by 25%-30% compared to conventional IDS.
To improve the consistency between color semantic expression and user emotional needs in graphic design, a graphic color matching method based on emotion-driven and digital visualization color mapping is proposed. This method quantifies users' ratings of pleasure and activation through a self-assessment model to form a two-dimensional emotional space. The K-means algorithm extracts color features from color palettes. A multi-task lasso regression model is used to jointly train an emotion prediction model. The L1/L2 mixed regularization is introduced to screen key features. In the optimization phase, multi-objective functions are used to synchronously optimize emotional bias and color distinguishability. A color mapping algorithm based on color gamut boundary intersection points is designed to reduce the cross-device color offset rate. Experiments showed that the model had high reliability in predicting pleasure and activation, with Cronbach's alpha reaching 0.9037 and 0.9045, respectively, and mean square error converging to 0.012. User testing showed that the optimized color palette's emotional coordinates reduced the deviation from the target emotion by 42%. In scenes such as film and television stylization and data visualization, the efficiency of color style conversion was increased by 35%. The color palette achieved an emotional matching accuracy of 89% while maintaining color harmony. The study addresses the limitations of traditional discrete emotion mapping, providing theoretical support for interdisciplinary research in color psychology and computer graphics.
In this paper, a novel tool for decision-making problems, namely the distance measure between intuitionistic fuzzy sets (IFSs) is being forwarded. Classically, the previous distance measures did not consider the impact of the hesitancy parameter, which produces counterintuitive outcomes when measuring distance. The proposed distance measure fully considers the influence of hesitation degree on distance measurement. Simultaneously, it is not only consistent with the four properties of distance measure, but also has nonlinear characteristics. The incorporation of the logarithmic transformation introduces nonlinear characteristics in the distance computation, which allows the measure to capture subtle variations between intuitionistic fuzzy elements more effectively. This nonlinear behavior improves the discrimination capability of the distance measure when comparing IFSs. Several examples were analyzed to demonstrate that the suggested distance measure produces preponderance outcomes. In addition, the applicability of the suggested distance measure was displayed by a series of applications in pattern recognition.
Regulatory Technology (RegTech) has emerged as a transformative field aimed at enhancing the efficiency and precision of regulatory compliance processes across industries. Originating from the financial sector, it leverages advanced technologies to address challenges in monitoring, reporting, and compliance. This paper explores state-of-the-art AI-based methodologies for compliance checking and proposes a comprehensive Compliance Management Framework (CMF) that integrates tasks such as compliance modeling, mapping, enactment, and assessment. Central to this study is the introduction of a unified Regulatory Compliance (RC) checking workflow, comprising four key stages: rule interpretation, rule mapping, rule execution, and compliance reporting. Emphasizing the complexity of regulatory compliance, the study identifies critical challenges, including computational infeasibility, semantic variability, and the dynamic nature of regulations. State-of-the-art compliance approaches are categorized into logic-based, NLP, ML, and DL techniques, with their application demonstrated in real-world contexts such as GDPR compliance and Building Information Modeling (BIM). The findings highlight the role of advanced technologies in improving compliance efficiency, accuracy, and cost-effectiveness, while identifying remaining challenges and opportunities for innovation. Methodologically, the proposed unified RC checking workflow offers an AI-driven, scalable compliance pipeline integrating rule interpretation, mapping, execution, and reporting via logic-based reasoning, NLP, ML, and DL. The proposed CMF introduces a unified compliance perspective by integrating compliance modeling, mapping, execution, assessment, and reporting within a continuous operational lifecycle. This integration enables a process-centric view that connects AI-based techniques with practical compliance operations, thereby extending existing RegTech survey studies toward a more structured methodological framework.
