https://repository.hneu.edu.ua/handle/123456789/140 2026-05-16T22:28:49Z 2026-05-16T22:28:49Z Minukhin S. https://repository.hneu.edu.ua/handle/123456789/40027 2026-05-14T21:21:50Z 2026-01-01T00:00:00Z

Назва: Development and Research of Batch Implementation of SQL-queries Based on the Rules of Their Ordering in Cloud Environments Автори: Minukhin S. Короткий огляд (реферат): The study proposes strategies for grouping (batching) queries in relational databases based on random grouping, as well as on prioritizing the values of individual performance metrics – execution time, DTU usage, and CPU load – and analyses their impact on the performance of the Azure SQL Database cloud platform service. The research methodology included creating a database in Azure SQL Database at different Azure service tiers – from S0 to S12 – to model various configurations of computing resources. To simulate realistic scenarios of working with the service, a database of a trading company with large sets of test data and several test database queries of varying complexity was used. Query batching strategies were developed: random grouping, grouping by ascending/descending query execution time, resource intensity (DTU consumption), and CPU load. Each strategy was tested across all resource configurations through multiple test trials, ensuring the relevance of the obtained results for an objective analysis. The results obtained demonstrated the necessity of using a differentiated approach to selecting query batching strategies depending on database size, query complexity, and the choice of query prioritization models in batch mode.

2026-01-01T00:00:00Z Shaposhnyk M. Minukhin S. https://repository.hneu.edu.ua/handle/123456789/40026 2026-05-14T21:17:59Z 2026-01-01T00:00:00Z

Назва: Mutual Information Preference Optimization for Robust Multi- Modal Recipe Generation Автори: Shaposhnyk M.; Minukhin S. Короткий огляд (реферат): This study evaluates the impact of Mutual Information Preference Optimization (MIPO) as a corrective layer within a hybrid vision-language architecture. Rather than introducing a new standalone framework, the research modifies an existing multimodal pipeline by integrating MIPO to bridge the operational gap between a DenseNet-121 ensemble and Llama 3.1 8B. The central hypothesis—that LLMs can act as autonomous semantic filters—was tested through contrastive alignment, which synchronizes CNN-derived visual features with the textual latent space. Experimental results on the Food-101 dataset validate this modification, demonstrating that the system can successfully suppress false-positive detections without a complete retraining of the visual backbone. By filtering out incongruous artifacts through preference optimization, the modified architecture achieved a 60,8% reduction in semantic hallucinations. This confirms the viability of using LLMs for real-time error correction in specialized domains, such as personalized dietetics, where output fidelity is a critical requirement.

2026-01-01T00:00:00Z Skorin Yu. Lukyanchuk S. https://repository.hneu.edu.ua/handle/123456789/39976 2026-05-12T20:31:23Z 2026-01-01T00:00:00Z

Назва: Artificial intelligence in automated financial risk management systems Автори: Skorin Yu.; Lukyanchuk S. Короткий огляд (реферат): The article is aimed at studying the use of artificial in-telligence in automated financial risk management systems to improve the accuracy, efficiency and efficiency of managerial decision-making in the financial sector. The study consists in a comprehensive study of the theoretical and methodological foundations of the use of artificial intelligence in management systems, analysis of modern approaches to the classification and assessment of financial risks using machine learning algorithms, formation of the architecture of the decision support system based on forecasting models, as well as assessment of the effectiveness of the built models based on the results of simulation modeling. Within the framework of the study, a model for forecasting the credit risk of bank customers has been developed, which allows assessing solvency based on historical data and modern machine learning methods. The research method is modeling using machine learning tools, including neural networks and ensemble learning methods (Random Forest), as well as data analysis using platforms to visualize results and evaluate the effectiveness of models. Particular attention is paid to data preparation, selection of relevant features, evaluation of model accuracy, and construction of interpreted visualizations such as SHAP graphs, ROC curves, etc. The result of the study was the creation of an effective model for predicting credit risk, which demonstrates a sufficiently high level of classification accuracy and the ability to adapt to changes in incoming conditions. The practical significance of the study lies in the possibility of implementing the developed model into the existing automated financial risk management systems of banking institutions, which will reduce the level of credit losses, increase financial stability and provide more accurate risk management. This approach contributes to the de-velopment of intelligent financial systems, increasing the level of automation of managerial decisions and strengthening the competitiveness of financial institutions in modern market conditions.

2026-01-01T00:00:00Z Minukhin S. V. Bukhalo V. O. https://repository.hneu.edu.ua/handle/123456789/39859 2026-05-07T18:15:25Z 2026-01-01T00:00:00Z

Назва: A study of the impact of the weighted reciprocal rank fusion method on the quality of recommender systems Автори: Minukhin S. V.; Bukhalo V. O. Короткий огляд (реферат): Providing high-quality recommendations is an important factor in increasing the level of audience engagement, since under conditions of rapid growth in the volume of available information, a significant part of which is presented in the form of implicit feedback, recommender systems ensure the effective selection and ranking of items according to individual user preferences. CF is one of the most widespread strategies for constructing such systems and generates recommendations based on the analysis of user interactions with similar behaviour patterns, which makes it possible to identify not only obvious but also unexpected potentially relevant items. During the generation of recommendations by different CF algorithms, recommendation lists for each user are produced that differ in the composition and order of items, which is caused by differences in the principles of determining the relevance of items. Since CF algorithms produce distinct recommendation lists for each user, it is appropriate to apply the WRRF method, which ensures the aggregation of rankings generated by algorithms in order to construct a single ranked recommendation list of higher quality. The purpose of this work is to study the influence of the WRRF method on the quality of generation and ranking of recommendation lists obtained as a result of the pairwise combination of CF algorithms through rank aggregation of items. According to the results of the experimental study, it has been established that the use of the WRRF method in the vast majority of cases ensures an improvement in recommendation quality compared with the best algorithm in the corresponding pair. The experimental evaluation was carried out using six CF algorithms on three datasets transformed into the implicit feedback format. The obtained results can be used in the development and improvement of industrial recommender systems in order to increase the quality of recommendation ranking without significant complication of their software architecture.

2026-01-01T00:00:00Z