In recent years developments in deep learning and machine learning techniques have had a major impact on the healthcare sector especially in the early diagnosis of diseases. The MedMNIST dataset contains medical images used for various diseaes diagnosis, and this dataset is analyzed with deep learning methods, making it possible to make important medical decisions. However privacy and security concerns encountered during the processing of medical data restrict traditional data processing methods. Federated Learning ensures data security by allowing training to be done on local devices without data being collected on a central server. In this study a method that improves the medical image classification task is proposed by integrating the federated learning approach with the Particle Swarm Optimization(PSO) algorithm. The PSO algorithm is used to optimize model parameters contributing to increased classification accuracy. In the proposed method the Convolutional Neural Network(CNN) model is used for image classification and local training is performed by preserving the privacy of the data with the federated learning method. Experimental results demonstrate that the proposed method effectively addresses the challenges of federated learning by achieving accuracy levels comparable to centralized state-of-the-art approaches, despite being trained across distributed clients. Furthermore, the method enhances data privacy by ensuring that raw data remains local, mitigating the security risks inherent in conventional centralized training paradigms. The integration of the PSO algorithm with the federated learning method allows the model parameters to be optimized more efficiently, increasing accuracy and making the training process more efficient. This method provides a safe and effective solution for medical image classification and has great potential especially in cases where data privacy is critical.
The dynamics of consumption of various medical products by the world's population and high competition in the pharmaceutical market make it urgent for pharmaceutical chains to develop new marketing strategies to improve the loyalty of consumers of pharmaceutical products. The purpose of this work was to study the consumer preferences of customers of a selected pharmacy chain in the Ukrainian city of Lviv, considering the vagueness of the primary information about these preferences obtained during their survey and using the Matlab computer system to perform numerical analysis of this information. The authors identify the primary factors that characterize the loyalty of consumers of goods and services in the pharmaceutical industry and present the results of a survey based on the developed questionnaire of customers of the pharmacy chain Biomed. It is proposed to apply the theory of fuzzy sets and fuzzy logic to calculate the integral indicator "Consumer loyalty to the pharmacy chain". The paper builds a fuzzy model for assessing consumer behaviour and loyalty using the Fuzzy Logic Toolbox module of the MATLAB application package, which allows, based on data about the consumer, to calculate their degree of loyalty to the goods and services of the pharmacy chain Biomed. Based on the research conducted, the average overall level of customer loyalty is estimated, and a system of measures is proposed to increase consumer loyalty, which will contribute to improving the company's marketing decisions and increasing customer satisfaction with the pharmacy chain.
In this study, we leverage the rich club coefficient to improve the representation capabilities of existing Graph Neural Networks(GNNs) for molecular graph classification. To achieve this, we introduce a novel structural plugin that enhances node features, enabling them to capture the hierarchical structural information within the graph. Our experiments demonstrate that our solution can uplift the performance of existing GNNs.
For many problems of modeling the behavior of complex systems, it is necessary to apply optimization procedures. The most effective in this case are stochastic optimization methods. Given the significant number of such methods, the problem of studying their effectiveness on the class of optimization problems arises. The complexity of such research is due to the need to automate the collection of statistical data on the relevant indicators of the effectiveness of a particular method. Obviously, it is worth developing appropriate software for this.This article considers the stochastic method of the guiding cone. Its main and additional parameters are determined, which directly or indirectly affect the effectiveness of the application of this method. An approach to software development is proposed, with the help of which it is possible to study the effectiveness of the method during the parametric identification of discrete dynamic models. Using the developed software, a numerical experiment was conducted to test the computational complexity of the method.General criteria for the effectiveness of optimization methods are determined. A procedure for automating the comprehensive testing of the effectiveness of various optimization algorithms within one stochastic method with changes in various search parameters is proposed. At the same time, the conditions for their application are indicated. The general characteristics of the software being developed are determined, and the choice of the software environment for it is justified. The concept of interaction of the developed software with a complex of computer modeling of complex systems is proposed.For the numerical experiment, a test objective function with a complex functional relief was selected in order to confirm the high efficiency of the selected optimization method. Based on the results of the numerical experiment, conclusions were drawn regarding the conditions for the effectiveness of the application of individual algorithms for solving a specific optimization problem.Directions for further research were formulated.
The work is devoted to the study of the economic activity of enterprises in the agrarian sector of the national economy, since this industry is developing and provides the largest share of exports in the national economy, even in the crisis year of 2022 for Ukraine. The possibility of applying modern scientific approaches to the modeling of leading agricultural companies using the cluster analysis method, and the implementation of the conducted modeling in modern information systems such as Python, which allows processing large data sets if necessary, and presenting them in a convenient interface for users, are considered. According to the results of the research, the proposed algorithm for clustering the leading agricultural firms of the national economy, and the result of the modeling is the construction of a dendrogram, which determines the rating position of the studied economic entities.
This study proposes a unified approach to developing individual and ensemble machine learning models for classification tasks in medical and financial domains, focusing on autism prediction and credit risk assessment. Using the Autistic Spectrum Disorder Screening Data for Adult and Home Equity Line of Credit (HELOC) datasets, six individual classifiers (KNN, SVM, Random Forest, XGBoost, MLP, Logistic Regression) were optimized and combined into ensemble models via stacking with hard voting, soft voting, and soft voting with Gompertz fuzzy ranking. The ensemble approach achieved superior performance, the best ensemble for autism prediction showed values of accuracy = 0.875 and F1-score = 0.873, and for HELOC, an accuracy = 0.735 and F1-score = 0.735. The results demonstrate the versatility and effectiveness of a domain-invariant methodology for classifying heterogeneous data, offering improved accuracy and robustness using ensemble techniques.
A nonlinear online bagging method is developed for integrating multiple systems in real-time data processing tasks. This approach utilizes a double neo-fuzzy neuron metamodel, which enhances approximation capabilities for combining systems outputs effectively. The training algorithm ensures efficient parameter tuning and supports real-time learning (online mode). Testing on short-term electric load forecasting demonstrated the effectiveness of this method, confirming its value in improving prediction accuracy through ensembling.
The secure and scalable key management of cryptocurrency custodians and exchanges is a significant challenge. Hierarchical deterministic wallets, introduced in BIP32 (ECDSA) and SLIP10 (EdDSA), provide structured key derivation, enabling secure multi-account management and improved scalability. However, software-based hierarchical deterministic wallets expose master keys to malware, insider threats, and memory extraction attacks. In this research, we propose an Hardware Security Module (HSM) secured hierarchical deterministic wallet architecture that integrates HSMs to ensure tamper-resistant key storage and enforce access controls. We analyze the security and performance of BIP32 versus SLIP10 in an HSM environment by measuring key derivation speed, transaction signing efficiency, and resistance to attacks. Analyses results show that SLIP10 offers stronger security guarantees, while BIP32 remains dominant due to blockchain compatibility. Additionally, the proposed architecture provides a more secure and scalable option for the key management of cryptocurrency custodians and exchanges.
This paper presents an improved McEliece cryptosystem based on the Redundant Residue Number System (RRNS). The proposed McEliece-RRNS scheme leverages the properties of RRNS–particularly efficient error correction and modular structures–to enhance cryptographic strength and reduce key sizes. The essence of the proposed McEliece-RRNS scheme is that, for public key generation, a generator matrix based on the redundant residue number system (RRNS) is used, in contrast to the classical McEliece cryptosystem, which employs a generator matrix based on Goppa codes. Compared to the classical implementation of McEliece, the system achieves smaller public and private key sizes, making it suitable for use in resource-constrained environments.
This paper proposes a five-stage framework for managing Agile software requirements using generative large language models (LLMs) and graph-based knowledge representation. A middleware system extracts Jira data, processes it via LLMs, and stores structured insights in a Neo4j graph. Evaluated on 18,199 real project tickets, the system achieved over 90% accuracy in user queries and demonstrated notable time savings. While initial LLM processing incurs higher costs, long-term use becomes more efficient especially with models like Gemini 1.5 Pro or other cost-redusing LLM solutions. The approach shows promise for scalable, intelligent requirements management and sets the stage for further validation and usability studies.
The EU Artificial Intelligence Act, implemented in 2024, was the first comprehensive regulation at the European Union level to concern the governance of artificial intelligence. However, despite its progressive approach, the aforementioned Act does not provide adequate protection against risks related to the use of generative AI systems for spreading disinformation, manipulating public opinion, and undermining democratic processes. This article analyzes the key gaps in the regulation, such as the absence of requirements for labeling AI-generated content, the lack of mechanisms for data verification, and insufficient legal accountability for human rights violations resulting from AI use. The author suggests the consideration of these issues in the context of other EU regulations, such as the Digital Services Act, and offers specific recommendations to improve regulation in these areas. The article includes a comparative analysis of different approaches to AI regulation in EU countries, including Spain, Lithuania, the Netherlands, Poland, and Germany, and provides recommendations for enhancing legal mechanisms to effectively counter disinformation and protect human rights.
This study explores the evolving concept of the immersive university, aiming to address the growing gap between the technological adoption of immersive tools and their strategic, pedagogically grounded integration within higher education institutions. The study also theoretically models an immersive university structure incorporating interdisciplinary collaboration, real-world engagement, and educational innovation. The analysis revealed a significant shift from isolated uses of immersive technologies to a more integrated and systemic understanding of their role within university ecosystems. Bibliometric mapping highlighted the increasing convergence of domains such as virtual reality, pedagogy, community engagement, and educational marketing. Additionally, the findings indicated a growing emphasis on scenario-based application of immersive tools, driven by pedagogical intent rather than technological novelty. The results suggest universities must adopt a holistic framework to effectively implement immersive strategies, including dedicated institutional units, ethical oversight, and stakeholder co-creation. The immersive university thus emerges not as a technological trend but as a strategic and structural innovation that enhances learning, fosters societal engagement, and redefines the institutional identity of higher education in the digital age.
The paper describes the results of the first attempts to use the Generative Artificial Intelligence (GAI) for High-Level Design (HLD) of Computer Devices based on the methods proposed and implemented by the author in some HLD tools. Mostly, Gemini was used as a GAI-assisted tool. The design process was organized in the following way: implemented algorithm description by High-Level Language (HLL), converting the HLL program to the three-address code, loops unrolling, three-address code parallelizing, and flow graph building, algorithm representation by the structural matrix, and computer device VHDL model building. Gemini implemented all steps of the design process. At every step, detailed instructions developed by author were given to Gemini, specifying the methods of activity at this step. For the design flow testing, the 64-point Radix-2 Decimation-In-Time (DIT) Fast Fourier Transform (FFT) algorithm was used. As a result, the FFT processor VHDL model was created and tested by GAI.
This study explores AI's transformative potential in higher education, focusing on Google Workspace for Education and Google's Gemini AI. A mixed-methods approach, including a 2021-2025 literature review, a VSPU case study, and a survey of 50 experts, identifies four AI trends: virtual tutors, automated assessment, content generation, and hybrid learning support. These enhance teaching efficiency, accessibility, and engagement. AI-generated lesson plans reduced faculty preparation time by 15 hours weekly, and transcription features increased non-native speaker participation by 30%. The VSPU case study shows reduced faculty workload and improved inclusivity. Challenges include digital inequality, teacher preparedness, and ethical concerns (data privacy, bias). Recommendations propose ethical AI policies, digital infrastructure investment, and teacher training for inclusive education. The study identifies four pivotal AI trends shaping the future of education: virtual tutors, automated assessment, content generation, and hybrid learning support. Virtual tutors, such as Gemini, deliver personalized learning experiences through adaptive explanations and real-time feedback. Automated assessment streamlines grading, reducing administrative burdens for educators. Content generation enables the creation of tailored educational materials, such as lesson plans and quizzes, enhancing resource accessibility. Hybrid learning support integrates AI into online and offline environments, promoting inclusivity through real-time transcription and translation features. These trends collectively enhance teaching efficiency, educational accessibility, and student engagement. Practical findings from the VSPU case study validate the benefits of AI integration. The study involved 30 pre-service teachers and 5 faculty members utilizing AI-enhanced Google Workspace tools, including Google Classroom, Docs, and Meet. AI-generated lesson plans in Google Slides reduced faculty preparation time by 15 hours per week, allowing greater focus on student interaction. Feedback in Google Docs improved group assignment quality, with 85% of students reporting clearer guidance. Analytics in Google Classroom identified at-risk students, increasing course completion rates by 10%. Transcription and translation in Google Meet boosted participation among non-native speakers by 30%, underscoring AI’s inclusivity. A survey of VSPU experts revealed that 80% consider AI essential for personalized education, yet 60% expressed concerns about ethical issues, including data privacy and algorithmic bias. Challenges encompass digital inequality (45% of educators reported technology access issues), inadequate teacher preparedness, and ethical risks, such as potential bias in assessments. To address these, the study recommends developing ethical AI policies, investing in digital infrastructure, and implementing teacher training programs. The research confirms AI’s transformative role while offering practical recommendations for leveraging Google Workspace and Gemini to foster inclusive and efficient educational ecosystems.
This paper presents an approach for implementing a GraphQL-based interface for querying CSAF (Common Security Advisory Framework) documents. The intermediate layer translates compact consumer queries into fully expanded GraphQL requests, handling nesting, enumeration harmonization, and logical operator alignment. It ensures complete document delivery, even when the consumer provides only partial structure definitions. Authentication via tokens restricts access based on TLP (Traffic Light Protocol) levels, enforcing secure distribution of CSAF advisories. The system is designed to minimize client-side effort while maintaining full data fidelity and security compliance.
An approach to identifying the coordinates of impulse sources (e.g., impurities) in space is developed. The time value during which a given substance passes through a certain point is taken as a priori information. The corresponding algorithm is constructed for the first time on the basis of a spatial generalization of numerical methods for solving boundary value problems on conformal mappings for domains enclosed by streamlines and equipotential curves. The results of numerical experiments are presented and analyzed.
The paper considers the decision-making system based on methods of teaching with reinforcement. The object of the study was the price range of companies in the stock market. The subject of research is the models of artificial intelligence based on reinforcement learning for the analysis of a stock market. This paper presents a reinforcement learning-based decision-making system for stock trading, incorporating an ensemble of three algorithms and trained on historical data from thirty companies. The system showed profitable performance in backtesting, proving effective in both stable and volatile market conditions. The results of the study are of practical importance not only for the financial sector, but also for the hospitality industry (HoReCa), where the use of artificial intelligence and predictive models allows optimizing pricing, improving revenue management, and increasing the efficiency of strategic planning.
The vibrations of unbalanced and balanced gasturbine engines grounding on their model as periodically nonstationary random process (PNRP) are analyzed. Low-frequency range ( <2 kHz ) of vibration signals was used to extract and evaluate of their regular component. The hidden periodical components of the first order in the vibration signal are detected using the least square (LS) method. Frequencies of the basic harmonics of deterministic component are estimated. Deterministic oscillations are found and compared for the cases of engines with balanced and unbalanced rotor. It is shown that the power of first harmonic for the unbalanced engine is dominant. The time dependences of PNRP mean functions using interpolation formula for balanced and unbalanced engines are obtained. Indicator, based on hidden periodicity harmonics ratio for evaluation of vibration condition of engine is proposed.
Modern transportation system gives specific requirements for performance of on-board navigation system to guarantee the safety of transportation. Integration of autonomous vehicles in transportation systems requires development of new positioning algorithms that will be computationally cheap for on-board equipment. Time Difference of Arrival (TDoA) positioning method grounds on measuring the time difference between two received radio signals from the network of synchronized ground transmitters. Time difference is proportional to the range difference which identifies a hyperbolic line of vehicle location. Navigation equation of TDoA is not linear and requires the usage of numerical methods to obtain vehicle position in case of multiple input data. In the paper, a TDoA position method has been implemented in the space of spatial indexes. Spatial index system partitions the shape of WGS-84 ellipsoidal model into a set of cells with hierarchical addresses. We use a regular hexagonal cell shape which has multiple benefits for navigation. Application of TDoA in hexagonal hierarchical spatial index provides a solution to navigation equation with the help of logical functions only. Logical operations and functions of set theory provide computational cheap identification of vehicle position that is important for modern autonomous vehicle development. In the numerical demonstration, a part of ground network of perspective eLoran system has been used for positioning with spatial indexes.
This paper explores the application of Generative Pre-trained Transformer (GPT) models in educational assessment. While traditional evaluation methods remain widely used, they face limitations in scalability, efficiency, and adaptability. GPT offers the ability to generate test questions, assess open-ended responses, and deliver adaptive feedback in real time. The study examines both technical aspects—such as system architecture, API integration, and data management—and pedagogical considerations, including rubric-based grading and alignment with learning outcomes. Key challenges, such as accuracy, bias, privacy, and academic integrity, are addressed. Results indicate that GPT-powered assessment systems, when combined with human oversight, can significantly improve efficiency and personalization while maintaining fairness and reliability.
