Social media platforms generate large volumes of health-related content that can provide valuable insights into emerging disease concerns and public perceptions. This study explores disease-risk trends through topic modeling and sentiment analysis of social media discussions. We used Latent Dirichlet Allocation (LDA) to extract topics and VADER sentiments from 50,000 statements related to diseases. The model performance was evaluated in terms of coherence, topic diversity, and stability metrics. The comparison was made with NMF and BERTopic. Three main topics were identified: COVID-19 symptoms and diagnosis; chronic disease management; and disease risk factors. The LDA model obtained coherence and topic diversity scores of 0.402 and 0.77, respectively. The sentiment analysis shows that the discussions are overall neutral, with some sentiments related to negative concerns. The results indicate that the use of topic modeling along with sentiment analysis is a useful approach to extract major disease-related topics from social media data. Future studies using transformer-based models, multilingual data, and longitudinal studies may improve digital disease surveillance and enable proactive public health monitoring.
Round-trip translation (RTT) is widely used for data augmentation and paraphrasing, yet its impact on meaning preservation remains poorly understood. We evaluate RTT systematically across two phases totaling 80,000 translation operations: 200 synthetic English passages are routed through single- and multi-hop pipelines spanning Bulgarian, Russian, and Chinese, translated by four systems ranging from a classical web translator (googletrans) to recent large language models (4o-mini, Claude Haiku 4.5, GPT-5-nano), with every passage undergoing ten RTT passes scored by BLEU, COMET-Kiwi, and xCOMET, joined by SBERT in the second phase. The results show that RTT outputs remain fluent while rapidly losing meaning: even the best configuration reaches only 0.646 xCOMET, and the four systems follow three distinct degradation regimes (front-loaded damage, continuous drift, and deterministic fixedpoint convergence). High surface overlap masks this loss, creating a surface fidelity trap, and genre-level analysis suggests that creative and news texts are most vulnerable. Sample-level inspection further reveals that xCOMET over-penalizes legitimate paraphrastic variation while under-penalizing domain-critical substitutions in legal text, exposing a mismatch between translation-oriented metrics and paraphrase evaluation. These findings argue against RTT in meaning-sensitive workflows and call for paraphrase-specific evaluation metrics.
Deploying machine learning models across distributed environments requires robust architectural patterns that ensure both cross-platform interoperability and secure access control. While Representational State Transfer (REST) APIs are widely adopted for service exposure, existing implementations in cross-machine scenarios remain vulnerable to token spoofing and post-hoc administrative tampering of centralized authorization logs. Therefore, this study presents a secure, deployable architectural pattern that integrates localized machine learning workflows with a decentralized authentication boundary. An online course recommendation system using text preprocessing, Term Frequency–Inverse Document Frequency (TF–IDF) vectorization, and cosine similarity is implemented and utilized explicitly as a functional demonstration case to validate the framework. The core machine learning engine is exposed via a Flask-implemented RESTful service layer, while secure access control is enforced by embedding a Proof-of-Work (PoW) blockchain token authentication mechanism at the API boundary. The empirical results demonstrate that access to protected endpoints is granted only upon successful validation of user tokens against immutable on-chain ledger records, structurally rejecting unauthorized requests. Performance benchmarking indicates that while the decentralized ledger operations introduce an operational latency trade-off (a 1113 ms mining overhead and 2.45-second end-to-end authorized query response), the proposed pattern provides an effective, tamper-evident firewall for delivering secure microservices in distributed client–server environments.
The automation of requirements engineering using large language models offers significant potential for efficiency but struggles with hallucinations and a lack of domain-specific precision in highly regulated fields such as healthcare. While retrieval augmented generation (RAG) addresses some of these issues, standard vector-based retrieval often fails to capture complex semantic relationships required for strict compliance. This research assesses the effectiveness of automated functional requirement extraction by comparing two distinct retrieval architectures: a baseline vector-only RAG versus a knowledge graph-enhanced RAG with three distinct prompt strategies. We implemented an end-to-end automated extraction pipeline applied to a corpus of heterogeneous healthcare documents. The study constructed two separate knowledge bases to perform a comparative analysis. Under the experimental conditions (seven healthcare documents, three security-focused testing areas, and Claude3-Haiku at temperature zero), knowledge base curation was found to be a stronger determinant of extraction quality than retrieval architecture, though the relative impact may vary with corpus scale, domain, and base LLM. With a curated knowledge base, GraphRAG eliminated the systematic relevance collapse observed under standard preprocessing, achieved significant structural traceability validity in independent Neo4j cross-checking, and produced prompt-stable performance across three prompt strategies. Without rigorous preprocessing, graph augmentation amplified rather than mitigated retrieval noise
Software Requirements Specification plays an important role in the success of software development projects by providing clear, structured, and consistent documentation of requirements. However, traditional SRS processes are manual, time-consuming, ambiguous, and inconsistent. This paper proposes an AI-enabled software requirements specification framework using Retrieval-Augmented Generation. The method combines Large Language Models with domain-specific knowledge retrieval to generate structured, coherent, and contextual requirement specifications. The proposed approach integrates document management, semantic retrieval, and generative reasoning to improve requirements quality, reduce ambiguity and enhance traceability. Validation was conducted using a banking chatbot case study. Experimental evaluation showed that the proposed approach achieved 0.75 precision, 1.00 recall, 0.857 F1-score and 0.75 accuracy for Functional Requirement identification on a manually labelled 12-sample evaluation subset. However, the system did not correctly identify Non-Functional Requirements in the current evaluation, indicating the need for a refined FR/NFR classification strategy in future work. The results demonstrate that Retrieval-Augmented Generation provides a practical and scalable solution for modern intelligent requirements engineering systems.
Low-Power Wide-Area Networks (LPWANs) based on LoRa technology have gained prominence in Internet of Things (IoT) deployments due to their long-range communication capabilities and low power consumption. The physical layer performance of LoRa’s Chirp Spread Spectrum (CSS) modulation under varying signal-to-noise ratio (SNR) conditions is critical for network planning and optimization. This study presents a comprehensive bit error rate (BER) characterization of LoRa CSS modulation across spreading factors SF7 through SF12 using Simulink-based Monte Carlo simulations across multiple channel conditions. The simulation framework implements a LoRa-compatible CSS physical layer model, capturing the essential spreading factor and processing gain characteristics without emulating proprietary chipset-specific implementations. Results demonstrate progressive BER improvement with increasing spreading factors under AWGN, Rayleigh fading, and Rician fading channel conditions, achieving error-free transmission at SNR values ranging from –6 dB (SF7 in AWGN) to –18 dB (SF12 in AWGN). The measured SNR sensitivity gain of approximately 2–4 dB per spreading factor increment confirms theoretical predictions and validates the simulation methodology. Higher spreading factors demonstrate superior resilience under fading conditions, with SF12 maintaining near-zero BER under Rayleigh fading at accessible SNR levels while SF7 does not. These findings provide baseline performance metrics essential for LoRaWAN network deployment planning, link budget calculations, and spreading factor selection in real-world IoT applications with diverse propagation environments. First published online 3 June 2026
Requirements engineering is essential for aligning software systems with stakeholder needs, yet incorporating stakeholder feedback into existing software requirements remains challenging due to ambiguity, inconsistency, and limited validation mechanisms. This paper proposes an AI-based evaluation approach for assessing the quality of software requirements and the correctness of feedback-driven modifications. The method combines natural language processing, large language models, and a hierarchical fuzzy inference system to support systematic evaluation. A multi-label text classification model is trained to identify common requirement defects, including ambiguity, subjectivity, vagueness, non-verifiability, and negativity. These indicators are aggregated using a fuzzy inference hierarchy to compute requirement quality scores and evaluate the preservation and correctness of changes introduced by stakeholder feedback. The approach is implemented as a web-based software system and evaluated against multiple AI-based methods and human assessments using defined quality criteria and performance metrics. Results indicate that the proposed hybrid method provides consistent and interpretable evaluations and can support more structured assessment of stakeholder feedback implementation in requirements engineering. First published online 3 June 2026
In modern organizations, the ability to efficiently manage and adapt business processes is essential. Business process modelling (BPM) is widely used to visualize, analyse, and improve operational processes. As the complexity of business environments increases, the integration of ontological modelling and data-driven approaches becomes increasingly relevant. Ontologies offer a semantic foundation for organizing and structuring process-related information, while data-driven methods support evidence-based decision-making and enable the adaptation of processes to dynamic conditions. Although both approaches show promise, the academic literature still lacks a coherent view of how they are jointly applied within BPM. This research conducts a rapid review of recent scientific publications to investigate how ontological and data-based methods are being used, what challenges are most often identified, and which research directions are emerging. The analysis reveals that the integration of these methods could address issues such as semantic consistency, process automation, and real-time decision-making. The results highlight existing research gaps and provide a clearer understanding of how BPM methodologies can be advanced by combining these two perspectives. This research contributes to the theoretical development of BPM by mapping current practices and offering insights for future researches. First published online 23 January 2026
The Analytic Hierarchy Process (AHP) is a human-centered method designed to structure complex problems and extract the authentic, consistent opinions of decision-makers. However, its practical application is often limited by inconsistency in human judgments, often caused by the respondent’s insufficient understanding of the task rather than simple mathematical error. The main goal of the article is to explore the possibilities of integration of innovative Artificial Intelligence (AI) tools for improving the AHP method. In order to improve respondent understanding and facilitate more intuitive and transparent consistency adjustments, this study also analyzes how to reduce the occurrence of inconsistency in pairwise comparison matrices and improve the Consistency Ratio (CR) by using advanced capabilities of large language models. The initial stage of study included a literature review, identifying typical problems in this area, reviewing the tools and methods used for obtaining better results, and presenting areas for improvement. At the second stage, the possibilities for improving consistency by increasing the influence of humans as decision makers, moving from the use of powerful mathematical optimization mechanisms to the application of human-centered explanatory AI techniques were analyzed. Based on the study results, the description of approaches for improvement of consistency in AHP was presented. First published online 27 January 2026
Video anomaly detection aims to identify unusual events in surveillance footage, yet many existing deep learning solutions remain too computationally heavy for real-time deployment on resource-limited hardware. This study presents a systematic comparison of three lightweight deep learning models for frame-level anomaly detection on the Avenue dataset, including a baseline 2D convolutional autoencoder, an enhanced reconstruction-based autoencoder with refined feature representation and decoding strategy, and a MobileNetV2-based supervised classifier fine-tuned for anomaly recognition. The baseline autoencoder achieves moderate detection performance, with an approximately AUC of 0.75. In contrast, the enhanced autoencoder improves reconstruction quality and raises the AUC to approximately 0.84 through more effective feature abstraction rather than increased architectural depth. The strongest results are obtained by the MobileNetV2 classifier, which achieves an AUC close to 0.99, high precision and recall, and a stable confusion matrix. These results demonstrate that lightweight architectures, when combined with appropriate training strategies and careful handling of class imbalance, can outperform more complex models. Overall, the study confirms that architectural efficiency and learning paradigm selection are more critical than model depth alone, making lightweight models well-suited to practical, real-time video anomaly detection scenarios. First published online 02 February 2026
Reinforcement learning (RL) algorithms, particularly deep reinforcement learning (DRL), have shown transformative potential in robotics by enabling adaptive behaviour in virtual environments. However, a comprehensive framework for efficiently testing, training, and deploying robots in these environments remains underexplored. This study introduces a standardized, open-source framework designed specifically for the Webots simulation environment. Supported by a robust methodology, the framework integrates innovative design patterns and the digital twin (DT) concept with three distinct design patterns for structuring agent-environment interaction, notably including a novel pattern aimed at improving sim-toreal transferability, to enhance RL workflows. The proposed framework is validated through experimental studies on both a model the inverted pendulum and a production-grade Pioneer 3-AT robotic platform. The experiments highlight the framework’s ability to bridge the gap between virtual training and real-world implementation. All resources, including the framework, methodology, and experimental configurations, are openly accessible on GitHub.
Fare prediction is a critical component of online ride-hailing services, as it significantly influences consumer decision-making and enhances operational efficiency for service providers. Reliable fare prediction is especially important in dynamic pricing environments, where fares are affected by factors such as demand fluctuations, traffic conditions, and weather patterns. This study aims to enhance fare prediction in ride-hailing services by utilizing advanced deep learning models. Using a comprehensive dataset of Uber and Lyft fare data collected in Boston during the winter of 2018, we evaluated three deep learning architectures: Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and BiLSTM with an attention mechanism (BiLSTM + Attention). The results showed that the BiLSTM + Attention model achieved the highest prediction accuracy, making it the most effective approach for fare prediction. However, its longer training time poses limitations for time-sensitive applications. Conversely, the LSTM model provided a strong balance between predictive accuracy and computational efficiency, making it a suitable alternative for scenarios that require faster model deployment. Additionally, our analysis identified key factors influencing fare variability – such as trip distance, time of day, and weather conditions – highlighting the importance of feature selection in enhancing model performance. By improving fare prediction accuracy, this study offers valuable insights for optimizing dynamic pricing strategies, enhancing consumer satisfaction, and helping ride-hailing platforms better manage supply–demand imbalances. These findings provide a foundation for future research exploring hybrid models and real-time data integration to further improve predictive capabilities in ride-hailing services.
In this study, a deep learning-based denoising autoencoder approach is proposed to increase the robustness of near-infrared spectroscopy data to random noise and improve quantitative modeling accuracy. Artificial Gaussian noise at four different levels (10, 15, 20, and 25 dB) was added to the near-infrared spectra obtained from milk samples to mimic the real measurement conditions. The noisy spectra were denoised by processing with an autoencoder architecture consisting of fully connected layers. The noise removal performance is quantitatively evaluated with both theoretical and measured signal-to-noise ratio values. The results show that the AE model significantly improves the spectral signal quality at all signal-to-noise ratio levels. In particular, at the lowest signal-to-noise ratio level (10 dB), the signal-to-noise ratio value nearly tripled to 29.6 dB with the autoencoder. At all other levels, an average increase of 18-20 dB was observed in the signal-to-noise ratio of the denoised spectra. In the second stage of the study, Partial Least Squares Regression models were built using both the noisy and cleaned spectra and evaluated on the test set with root mean square error and coefficient of determination. The Partial Least Squares Regression models built with the denoised spectra achieved lower root mean square error and higher coefficient of determination values at all signal-to-noise ratio levels. Especially at the 10 dB signal-to-noise ratio level, the coefficient of determination value of the model increased from 0.44 to 0.71, while the root means square error decreased from 0.60 to 0.43. The results show that the deep learning-based AE architecture can effectively reduce random noise in near-infrared spectral data and significantly improve both spectral signal quality and quantitative modeling performance. This approach provides an effective solution to improve model reliability and accuracy in near-infrared spectroscopy analysis.
The development and assessment of modern weapon systems require efficient and flexible simulation tools. This paper introduces a multi-scale discrete-event simulation framework designed to evaluate the dynamic combat effectiveness of weapon systems. The framework combines high-resolution and low-resolution models to address the complexities of real-world engagements while maintaining computational efficiency. Physical processes are encapsulated as modular state transition functions, allowing seamless integration of a multi complexity level modeling approach. The framework’s versatility is demonstrated through a case study analyzing the effectiveness of a tank weapon system against a fleet of drones. Non-deterministic methods such as Monte Carlo simulations for uncertainty quantification are used to evaluate probabilistic key metrics, such as projectile accuracy and lethality, providing insights into engagement dynamics and optimization of firing strategies. By leveraging a hybrid continuous/discrete approach and modular design, the framework enables comprehensive assessments of weapon effectiveness during an engagement, bridging gaps in traditional deterministic methodologies for both static and dynamic targets. Future enhancements will focus on optimizing sampling techniques for broader applicability of high-resolution stochastic simulations in modern combat scenarios.
The growing accessibility of satellite imagery and the rapid evolution of machine learning (ML) techniques have significantly advanced land use classification for environmental monitoring. However, challenges such as cloud coverage, varying image resolutions, and seasonal changes continue to hinder classification accuracy and consistency. This study aims to improve land use classification by proposing an integrated cloud interpolation, vegetation indices and ML based approach for classification of Sentinel-2 (S2) satellite data across the Baltic States. Specifically, a spatiotemporal interpolation module is introduced that reconstructs cloud-obscured pixels using multi-temporal coherence and derives optimized vegetation-index composites to enhance class separability under varying seasonal conditions. In order to achieve this aim and to choose the best ML algorithm for land use classification, we compare the performance of three classification algorithms, i.e., Random Forest (RF), K-Nearest Neighbours (KNN), and Support Vector Machines (SVM), and evaluate their effectiveness in handling noisy and incomplete data. Our experimental results show that all three methods achieve strong classification accuracy, with RF exceeding 90%, while KNN and SVM also demonstrate competitive results. These methodological enhancements have been demonstrated to reduce cloud-induced misclassification and provide a scalable, transferable framework for operational land-use mapping in challenging atmospheric and seasonal contexts. These findings highlight the robustness of the proposed approach and provide valuable insights for future applications of ML in land use classification and environmental analysis.
Artificial intelligence-based solutions are widely used in different areas. With the advent of chatGPT, education has faced many challenges, such as student cheating by generating text in various practical and assessment tasks. As the number of large language models increases, it becomes difficult to control their use, and their capabilities increase over time as well. However, large language models do not only provide a negative aspect, but when used properly, they can be applied to useful and meaningful solutions. One of these is the personalization of learning, which would help to direct learning to the right needs without much human intervention, for example, when there are certain difficulties, knowledge gaps or lack of motivation. This manuscript, using a systematic analysis of the scientific literature, reviews technological solutions that are already currently used in personalizing learning. It also reviews the latest trends and challenges that would allow this area to be raised to a higher level.
The rapid development of information and communication technologies (ICT), and enforcement of digital transformation have caused the need to develop and deliver software in even shorter time than previously, and has become crucial to businesses. Yet, this new need has caused a challenge of finding expert IT workforce able to deliver software applications due expected time. Moreover, in a rapidly changing business world, the traditional software development can no longer keep pace and provide solutions to businesses fast enough. Low-code development and inclusion of so-called citizen developers are seen as a solution for the problem. This paper explores the introduction of rapid application development courses on Oracle APEX technology to students of a European technological university. The paper discusses the implementation of the course and its design, and looks how the developed basic-level course was received by students. It is shown that the courses have been well-received by students, who based on provided feedback are satisfied with the online set up for self-regulated learning of rapid application development skills.
The use of artificial neural networks for the optimisation of factory layouts is not a common practice, primarily due to the challenge of collecting sufficient layout data to form datasets for artificial intelligence (AI) model training. This paper presents a supervised learning method derived from a PhD thesis that employs neural networks to assess factory layouts. The training data is generated using a random layout algorithm, which is capable of producing numerous layouts. These layouts are then labeled through a discrete event simulation. The combination of layouts and simulation metrics serves as the training basis for the neural network. The AI framework integrates a convolutional neural network with a multilayer perceptron, which is capable of handling both tabular and image data. Ultimately, this allows us to calculate of the simulated throughput.
The increase in software complexity, driven by technological developments and user demands, has created major challenges for companies in Software Quality Assurance. Companies seek efficient ways to identify and mitigate defects, recognizing that they cause high financial costs and other problems with negative impacts on business. Among defect prediction approaches, Just-In-Time Software Defect Prediction has received increased attention from software industry professionals in recent years. This technique aims to identify and treat defects early, to improve the quality of the software development cycle. This study proposes a Deep Learning-based approach for Just-In-Time Software Defect Prediction using a large dataset of historical data from several popular software projects. The Deep Learning model was trained to identify defects by analyzing the software metrics provided by the dataset. The model achieved an accuracy of 82.08% in its predictions, and it was possible to determine the most relevant metrics for its conclusions through interpretability techniques. The results obtained demonstrate the potential of Just-In-Time Software Defect Prediction as a tool for improving software quality and encouraging the development of new studies and improvements in this area of research.
Machine learning (ML) algorithms are more and more widely applied in various types of systems, so the research related to them is also increasing. One of the areas of research under consideration is the classification of non-functional requirements (NFRs) using ML algorithms. This area of research is important because the automatic classification of NFRs using high-performance ML algorithms and corresponding features helps requirements engineers classify non-functional requirements more accurately. This paper examines ML algorithms suitable for solving classification problems and their effectiveness in classifying non-functional requirements. Based on the described stages of the research methodology ML algorithms models were compared using the accuracy, precision, recall, and F-score metrics. A majority voting classifier model was created using Support Vector Machine, Naïve Bayes and K Nearest Neighbor Algorithm algorithms. After K-Fold cross validation were obtained these results: accuracy – 0.710 (scale from 0 to 1), precision – 0.845, recall – 0.814 and F-score – 0.815.
