Person Re-Identification (ReID) remains a pivotal challenge in intelligent surveillance, constantly tested by dramatic intra-class variations, pose changes, lighting shifts, occlusions and distracting backgrounds. This paper proposes a self-attention based hybrid framework (VSPeR) designed to improve data diversity and visual quality for person re-identification in real-world scenarios. At the heart of VSPeR lies a powerful synergy: the robust, identity preserving feature extraction of Variational Autoencoders (VAEs) meets the precise, context-aware focus of Self-Attention Generative Adversarial Networks (GANs). This hybrid architecture captures rich, semantically meaningful latent representations while zeroing in on the most discriminative visual cues necessary for accurate cross-view matching. The outcome is high-fidelity, identity-consistent synthetic images that significantly amplify the training dataset enabling smarter, more effective learning. Put to the test across four leading ReID benchmarks, VSPeR achieves competitive performance compared to existing approaches in terms of mean Average Precision (mAP) and Rank-1 accuracy.
In this study, EEGNet-CA, a lightweight deep learning model with an integrated attention mechanism, is proposed to diagnose Attention Deficit and Hyperactivity Disorder (ADHD) in children via EEG data. The proposed model is able to learn the contribution of each EEG channel in the classification and focus on important signals by integrating the channel attention (CA) mechanism into the classical EEGNet architecture. In this context, 19-channel EEG recordings from 121 children aged 7–12 years, obtained from the IEEE dataset, were used. The raw EEG signals were scaled with z-score normalization to improve the stability of the learning process and segmented using the 50 % overlap sliding window method. The EEGNet-CA model, developed using the obtained EEG segments, was utilised for classification purposes. Experimental results show that the proposed EEGNet-CA model achieved a 99.79 % F1-score and exhibited statistically significantly superior performance compared to standard EEGNet, 1D CNN, 2D CNN, and artificial neural network-based methods. In addition, high accuracy even with only frontal region channels supports the neurophysiological importance of this region in ADHD diagnosis. The results show that the CA mechanism is an effective method to increase the classification success in EEG-based diagnostic systems.
Recommendation systems built on ratings often fail to fully exploit the rich information contained in textual reviews and social-platform content. Text provides not only an additional signal for ranking, but also a natural basis for generating explanations. This study introduces XTRA, a rating-free, text-driven recommendation system that derives user relationships from shared review content and transforms sentiment expressed in text into personalized recommendation decisions. The proposed approach also incorporates explainability mechanisms to clarify both why users are considered similar and how textual sentiment contributes to the final recommendation. Experimental results on Amazon-Beauty, Yelp-Illinois, and IMDB-20K show that XTRA outperforms strong rating-based baselines on Yelp and IMDB, while remaining competitive on Amazon. These findings indicate that textual signals can serve as an effective alternative to ratings while also supporting transparent and interpretable recommendations. Implementation details are provided to support reproducibility and future extension.
This study investigates adaptive synthesis of intelligent monitoring (IM) for applications modified by large language models (LLMs). A multi-objective, property-driven genetic approach with a tensor of chromosome representation is proposed to generate monitoring configurations under architectural changes. Using migration from Oracle to PostgreSQL as a controlled case, the study shows that functional equivalence can be maintained while monitoring-related properties drift, supporting adaptive archi tecture-aware monitoring over static settings.
Fire and smoke detection in image data is a crucial application of computer vision; however, practical deployment in real-world environments remains challenging. While previous research has made significant progress; however, detection systems often struggle to maintain stability when faced with changing lighting conditions, partial occlusions, and the complex visual characteristics of indoor spaces. This study proposes an automated fire and smoke detection system based on the Real- Time Detection Transformer (RT-DETR) architecture. Unlike traditional models that focus solely on accuracy, this system is engineered to address the practical need for early warning, achieving a high recall of 91.6 % to minimise missed fire events. The system is designed for versatile integration into existing CCTV surveillance, Smart Building ecosystems, and IoT/Edge. Evaluated on the Home-Fire dataset using a five-fold cross-validation strategy, the model achieves an mAP@0.5 of 94.5 %. These results demonstrate that the proposed system offers a robust, scalable, and reliable solution for real-world fire safety monitoring, providing a robust foundation for autonomous fire safety monitoring and rapid emergency response.
The growth of Automated Machine Learning (AutoML) has expanded access to machine learning workflows by enabling the automation of tasks and reducing the technical barrier to entry. However, the reliability and maintainability of these libraries depend on the quality of their underlying source code. This study presents a novel, systematic analysis of 16 Python AutoML libraries utilising SonarQube - an industry-standard SCA platform - and Python analysis tools: Bandit, Coverage.py, Prospector, Pylint, Radon, and Ruff. The AutoML Libraries are evaluated using software quality metrics, which collectively reflect overall code complexity, maintainability, security, and adherence to Python coding standards.Strong agreement was observed between SonarQube-based rankings and rankings derived from Python-based tools. Based on median SCA rankings, the libraries were ordered (highest to lowest estimated code quality) as follows: Hyperopt-sklearn, AutoKeras, GAMA, MLBox, FEDOT, TPOT, MLJAR, LightAutoML, Auto-sklearn, PyCaret, FLAML, Auto-PyTorch, Ludwig, EvalML, AutoTS, and AutoGluon.An additional exploratory Spearman rank correlation analysis examined the relationship between SCA metrics and forecasting performance measures from a prior electricity price prediction benchmark (n = 7). Several SCA metrics exhibit strong monotonic relationships with forecasting error measures, e.g., SonarQube Violations and Code Smells correlate positively with mean absolute error (rho = 0.86), while Class Cyclomatic Complexity (rho = -0.89) and Duplicated Files (rho = - 0.86) correlate negatively with library execution time. Due to the limited sample size, these findings are descriptive and non-parametric. The results suggest that code quality scores may relate to lower-bound predictive performance and computational efficiency, warranting further validation.
Online reviews influence consumer decisions in e-commerce but are threatened by deceptive content. Detecting these fake reviews across domains remains challenging due to limited labelled data, class imbalance, and domain-specific bias. This work proposes CORAL, a unified retrieval-augmented framework for cross-domain fake review detection that mitigates source-domain bias through unified masking and generates targetaware counterfactuals via retrieval grounding with orthogonal alignment. Unlike prior domain-adaptation approaches, CORAL enables zero-shot classification without manual annotation. Evaluated across twelve cross-domain tasks on benchmark ecommerce datasets, CORAL achieved average gains of 21.74 % in accuracy and 24.25 % in F1-score over strong baselines while improving trustworthiness and robustness. It demonstrated stability across random seeds, resilience to moderate noise, reduced hallucination, and competitive runtime efficiency. Additionally, CORAL enabled the annotation of a novel healthcare review dataset, addressing scarcity of labelled data.
A linear price-based transportation model for energy allocation in decentralised energy markets with multiple prosumers and consumers is studied. The centralised formulation minimizes the aggregate difference between generation costs and local consumer prices under supply and demand constraints. To model limited information and reduced coordination, a parametric decentralised variant is introduced, where consumers are clustered into K groups, and prices are averaged within K clusters. Consumer clustering is based on price similarity. The proposed approach treats decentralisation as a tunable structural parameter rather than a binary rule. Computational experiments across varying market sizes demonstrate that clustered configurations frequently outperform the fully individualized pricing regime in terms of the objective value, especially in unfavourable market conditions. A finding is the efficiency-versus-decentralisation relationship is non-monotonic, with an instance-dependent optimal number of clusters. The results suggest that controlled price aggregation can improve market outcomes while reducing informational and computational complexity.
This study puts forward a Hamiltonian-inspired modification of Particle Swarm Optimisation (PSO) algorithm. Since the standard PSO procedure does not take into account physical properties like particle masses, geometrical sizes, and energy consumption, it is not fully applicable as a navigational and coordination tool in real-world environments. In particular, generic PSO mechanism cannot stop the particles from collisions. To address these issues, we propose a new PSO formulation based on a Hamiltonian interpretation. This approach allows bringing together the kinetic and potential energy terms with the forces acting on the agents, as well as the derivation of agent’s velocities and positions. The potential energy represents attraction toward both personal and global best positions in a spring-like manner. As a component of conservative forces derived from the potential energy term, we introduce a special repulsive potential function to prevent collisions among agents. The kinetic energy, which is derived via agent mass and momentum, determines the movement dynamics. To model the energy loss, we incorporate Rayleigh dissipation term that accounts for non-conservative forces. According to the proposed model, agent displacements are computed using the obtained velocity and momentum vectors. Additionally, we introduce individual and swarm energy efficiency metrics to study the agents’ motion in a 2D testing environment. The presented approach enables stable, coordinated, and collision-free multi-agent motion within a physics-inspired optimisation framework.
Efficiency, convenience, and safety are the standard promises of modern engineered systems. However, the execution often fails to match this vision, leading to a disconnect where users become skeptical, frustrated, or simply unwilling to adopt the technology in their daily lives. To bridge this gap between technical potential and user reality, this paper proposes the User-Aligned Systems Engineering Framework (UASEF). While the framework is deeply rooted in the complexities of smart home research, its core principles are designed to be universally applicable, reorienting the engineering process to place the human element at the centre. UASEF mandates a structure built around a central core of security and trust by design, revolving through six iterative phases, moving beyond basic functionality to prioritize deep stakeholder analysis, transparent architecture, and critical factors like cost, accessibility, and embedded security. By deconstructing specific friction points such as usability barriers and privacy concerns, this study demonstrates that the design principles required for a smart home are actually vital for any complex system. Ultimately, UASEF provides developers with actionable guidance to create technology that is not merely functional, but inherently secure, intuitive, and capable of earning long-term user confidence.
Chest X-ray lesion detection remains challenging due to severe class imbalance, subtle lesion appearance, and the risk of over-optimistic evaluation caused by improper data splitting. In this study, we propose a sensitivity-oriented detection framework based on YOLOv11 for robust chest X-ray screening under clinically realistic conditions. The proposed approach integrates patient-wise data partitioning, enhanced data augmentation, and prediction fusion to improve generalization while mitigating data leakage. Experiments are conducted on the VinDr-CXR dataset using a strict patient-level split to ensure full separation between training and validation sets. A series of internal fine-tuning scenarios is designed to analyse the trade-offs among precision, recall, and localization accuracy. Based on internal validation, the medium-scale YOLOv11-m configuration (denoted as M3) is selected as the reference model, as it provides the most stable balance between sensitivity and localization performance. Under rigorous evaluation, M3 achieves a precision of 0.431, a recall of 0.416, an mAP@0.5 of 0.387, and an mAP@0.5:0.95 of 0.193. Compared with representative baselines, M3 demonstrates improved robustness under patient-wise evaluation, outperforming transformer-based DETR by a large margin (mAP@0.5: 0.387 vs. 0.232) and achieving performance comparable to YOLOv7 while exhibiting substantially higher sensitivity to small and diffuse lesions. Further comparison with recent studies shows that the proposed method achieves higher overall mAP@0.5 (0.387 vs. 0.362-0.378) while improving detection performance on clinically challenging abnormality classes. These results indicate that the proposed YOLOv11-based framework provides a reliable and clinically meaningful baseline for chest X-ray lesion screening and future methodological advancements.
Digital twins increasingly rely on autonomous self-updating mechanisms to remain synchronized with physical systems; however, repeated self-updating can lead to error accumulation, numerical instability, and progressive loss of physical consistency when models iteratively learn from their own predictions. To address this challenge, the study proposes a physics-stabilized self-updating digital twin framework based on Physics-Informed Neural Networks (PINNs) and demonstrates its core principles on a canonical thermal field prediction problem. The framework integrates adaptive physics-loss weighting, a physics-only stabilization stage, and second-derivative smoothness regularization within the self-updating loop, enabling controlled data assimilation while explicitly enforcing governing equation constraints. Numerical results show a monotonic reduction in root mean square error (RMSE) from approximately 1 & times; 10-3 in the first update cycle to 8 & times; 10-5 after four update cycles, accompanied by effective suppression of model drift and a substantial reduction in partial differential equation (PDE) residuals compared to a na & iuml;ve self-updating strategy. Furthermore, the analysis reveals the existence of an update saturation point, beyond which additional autonomous updates yield diminishing accuracy improvements, providing a physically motivated stopping criterion for autonomous updating. By establishing a stable and physically interpretable self-updating architecture, the study provides a foundational framework for the development of reliable digital twins, with clear potential for extension to more complex thermal and Multiphysics systems.
Artificial intelligence (AI) and AI-based systems are rapidly gaining popularity across all areas of daily life. Among these systems, large language models (LLMs), which probabilistically model language to understand and generate text, stand out at the forefront. The ability to generate results from LLMs, whose primary focus is language, is of significant technical and social importance. As language diversity increases, the ability of LLMs to produce stable and consistent results is trending downwards. This decrease has a close relation with the size of the model, the scope of the training data, and the prompt technique used in response generation. To this end, a study was conducted to measure the success of LLMs in different languages. In the study, four LLMs were examined, three of which were open-source (DeepSeek-Coder-6.7B-Instruct, Qwen2.5-Coder-7B-Instruct, Llama-3.1-8B-Instruct) and one was closed-source (GPT-5). These models were evaluated using the HumanEval-XL dataset across seven natural languages that have different data sources and usage prevalences. Additionally, the effect of the human development index (HDI) values of the countries where the languages are spoken and the prompt technique used on the results was also analysed. Results show that as LLMs grow, performance differences between languages have decreased. Additionally, it has been observed that whether the models are open-source or closed-source also has a significant impact on performance. Among open-source LLMs, DeepSeek-Coder-6.7B-Instruct's accuracy rates range from 37 % to 60 %, while Qwen2.5-Coder-7B-Instruct and Llama-3.1-8B-Instruct have performed more consistently in the 95–99 % range. GPT-5, which is a closed-source LLM, has demonstrated balanced accuracy across all languages. The results obtained reveal remarkable results in ethics, quantity of linguistic data, and equality of access to technology. The results also clearly demonstrate the relationship between multilingual accuracy, language prevalence, and prompt techniques. In this way, the study offers a clearer and more comprehensive understanding of the issues surrounding linguistic justice and the generalization of LLMs in the field of AI.
The growing reliance on Electronic Document Management Systems (EDMSs) in public institutions necessitates an improved understanding of usability across diverse user groups. This study evaluates the perceived usability of a widely deployed commercial EDMS used in a public university context by examining differences across gender, age, personnel type (academic/administrative), education level, and prior EDMS experience. A mixed-methods design integrates objective task-based performance measures (completion, time, and perceived difficulty) and the System Usability Scale (SUS-TR) with qualitative feedback collected via open-ended questions. Overall, users valued the system’s contribution to streamlining document workflows; however, they reported notable usability barriers, particularly related to complexity and navigation. No significant differences were observed by gender, age category, or prior EDMS experience, whereas administrative staff and participants with higher education levels reported higher SUS scores. The mean SUS score (53.25) indicates below-average perceived usability when interpreted against established benchmarks, suggesting the need for targeted usability improvements. Qualitative feedback further highlights the need to simplify interaction flows – particularly around document search, dispatch, and leave management – and to enhance training and support resources to address recurring usability issues and reduce user errors. The study offers actionable recommendations for improving EDMS usability in public institutions and underscores the importance of user-centred design in digital transformation initiatives.
The growing problem of air pollution by fine particulate matter (PM2.5) from local sources, such as boiler houses or small industrial facilities, requires effective and accessible assessment tools. A significant gap exists between complex, resource-intensive dispersion models used in research and the practical needs of engineers, ecologists, and regulatory bodies who require instruments for rapid operational analysis. This problem is particularly acute in regions like Ukraine, where access to real-time, high-resolution environmental data is limited, and regulatory practices often rely on legacy methodologies. The paper describes the development and testing of a desktop software application with a graphical user interface (GUI) designed for scenario modelling (“what if” analysis) and quantitative assessment of air pollution levels from a local source. The core of the software tool is based on an adapted Gaussian plume analytical model, which calculates pollutant dispersion considering meteorological conditions and source parameters. The system integrates a developed method for integral impact assessment, categorising the pollution level based on calculated concentrations. The developed software allows the user to interactively input the constructive (stack height, diameter) and operational (emission rate) parameters of a pollution source, as well as current meteorological conditions. The system provides an instantaneous calculation of the expected PM2.5 concentration at a given point and classifies the impact: “Low”, “Moderate”, “High”, or “Very High”. The developed tool brings practical value, supporting the decision-making process. It provides a means for the operational monitoring of environmental impact and the preliminary planning of measures to reduce the ecological load from local pollution sources, making complex analysis accessible to a wider range of specialists, especially in data scarce environments.
The reliable and timely detection of cracks in concrete structures is essential for maintaining the safety, functionality, and longevity of civil infrastructure, including buildings, bridges, highways, and dams. Structural cracks can emerge due to multiple factors such as material fatigue, environmental stressors, seismic activity, and thermal expansion, necessitating accurate and efficient monitoring systems. Traditional inspection techniques, including manual visual inspection and non-destructive testing, are labour-intensive, prone to subjectivity, and often lack scalability. To address these limitations, the research presents CrackNet-VGG, a deep learning-based framework that leverages the VGG16 convolutional neural network architecture for automatic binary classification of surface cracks in concrete images. The proposed model leverages transfer learning by fine-tuning the VGG16 architecture on concrete surface datasets, utili sing its convolutional layers for robust feature extraction and its fully connected layers for final binary classification. The model is trained and evaluated on publicly available benchmark datasets, categorised into two classes: cracked and non-cracked surfaces. Experimental results demonstrate that CrackNet-VGG achieves a high classification accuracy of 96.07 %, with 95.57 % precision and 95.31 % recall, surpassing several baseline deep learning models in terms of accuracy. These results validate the applicability of CrackNet-VGG as an effective solution for automated concrete crack detection in real-world scenarios.
To improve the network interaction step-wise process, a genetic algorithm is suggested for finding more stable solutions in bimatrix games. The algorithm is based on using an approach of successive approximation to an equilibrium situation within a finite space of actions whose size directly depends on the number of game repetitions (network interactions). The algorithm has seven input parameters: the population size, the maximum number of generations, the number of generations for the early stop, the mutation rate, the number of bits per pure strategy, the number of maximum network interactions, and the number of the best chromosomes selected. The algorithm is more efficient for fewer network interactions, when the network peers obtain more stable and consistent strategies that encourage interaction itself rather than resigning from sending any information due to instability. The equilibrium concept is strengthened by introducing a criterion of mutual profitability, which is clearly a good tone and respect in network interactions (particularly, in P2P file-sharing networks). This criterion, expressed as a fitness value equal to negative maximum of potential losses, can be varied to alternatively evaluate the consequence of swerving from a given mixed strategy.
This work builds a benchmarking pipeline for resource deployment, configuration input, and the evaluation of Cassandra and ScyllaDB performance. It investigates performance under different scenarios, workload types, and internal structures. Insights and future improvements are provided. Analysing comparisons of both databases, there were several notable differences between Cassandra and ScyllaDB. ScyllaDB demonstrated superior performance in production-ready materialized views, global secondary indices, lightweight transactions, change data capture, row-based data cache, and adaptive behaviour to real-world workloads. On the other hand, Cassandra exhibited advantages, such as a well-established size-tiered compaction approach and the ability to leverage existing Java Virtual machine (JVM) tuning techniques.
Whisperer is introduced as an intelligent, real-time prompting system that aims to improve the flow and naturalness of speaking in public and on camera. It is different from regular teleprompters because it does not just follow a script. Instead, it uses Google Cloud’s low-latency speech-to-text (STT) and text-to-speech (TTS) services to sync spoken content with a prepared script in real time. The system can handle synonyms, homophones, numeric variations, and spontaneous improvisations because it uses linguistic models such as CMUDict for phoneme-level alignment, FastText for semantic similarity, and BERT for contextual understanding. Whisperer also has adaptive TTS feedback that matches the speaker’s speed. This includes changes made in real time based on how long the speaker pauses and how fast they speak. Testing shows that the speakers’ fluency and consistency of delivery have both improved.
Systematic collection of gait data directly affects the reliability of gait analysis. Therefore, the first step that should be given importance in gait analysis is to ensure that the data collected is of high quality and reliable. Optimising and standardising the data collection process is a critical requirement that increases the success of the analysis. This study proposes a systematic multi-camera placement strategy and data processing process to collect gait data in real time with RGB cameras for gait recognition and analysis applications. The proposed method provides an end-to-end framework from the physical setup of the data collection environment to the data processing steps. The camera placement strategy aims to maximise the visibility of all body parts by capturing the participant’s body from different directions during walking. The results of three different methods used for silhouette extraction from the acquired videos were compared. Furthermore, a video-based approach was used to calculate participants’ walking speeds. Theoretical and practical information provided regarding the data collection and analysis process is detailed to guide future studies. In this respect, the paper is aimed to be a guiding study for researchers working in the related field.
