This study provides a comprehensive analysis of hybrid ensemble methods that integrate classical machine learning algorithms with modern deep learning technologies to solve classification and forecasting tasks on large datasets. The main goal of this work is to develop and empirically validate a methodological approach that allows for achieving an optimal balance between model performance and the explainability of its decisions. The study used stacking, bagging, and boosting methods in combination with interpretable machine learning techniques, including SHAP analysis and feature importance methods. The results of the empirical study demonstrate that the proposed hybrid architecture improves classification accuracy by 12–18% compared to the baseline models, while maintaining an interpretability level above 0.85 using the LIME metric. It has been established that the optimal ensemble configuration includes a combination of random forest, gradient boosting, and neural networks with weight coefficients of 0.4, 0.35, and 0.25, respectively. The theoretical significance of the work lies in expanding the methodological framework of data mining by integrating the principles of explainable AI into ensemble architectures. The practical value is determined by the possibility of applying the developed approach in critical areas that require transparent decision-making.

The paper proposes a theoretically grounded and computationally efficient algorithm for identifying critically important (“bridge”) edges in multilayer social graphs. The approach consists of three sequential procedures: 1) spectral coarsening of each graph layer — compression while preserving key spectral properties (in particular, the Laplacian) and the local resistive structure; 2) approximate estimation of edge betweenness centrality — computing edge importance with layer weights and cross-layer interactions taken into account; 3) greedy shortest-path coverage — iteratively selecting edges that maximize graph fragmentation while minimizing the number of removals. The key properties of the algorithm are established: it preserves above-threshold connectivity of the residual graph (via control of the second Laplacian eigenvalue); limits the growth of the effective diameter after edge removals; provides a (1–1 / е) approximation to the optimal shortest-path cover; is robust to noise perturbations and variations in layer weights; and has asymptotic complexity O(Ln log n + km log n), which is substantially lower than that of classical methods. The practical significance lies in monitoring information diffusion, assessing structural vulnerability, and predicting cascading failures in multilayer infrastructures (social platforms as well as transportation and communication networks). Limitations include the assumption of shortest-path propagation, a priori layer aggregation, and the lack of an explicit temporal dynamics model.

Presentations have become an integral part of the core processes of many areas of professional activity. Nevertheless, the market for professional software packages that allow you to create presentations that meet modern requirements is not so saturated with competition. Microsoft’s software products are the undisputed leader. Competitors are much inferior in quality. But, as mentioned in previous articles in this series, cross-platform file formats are not protected from simple injections (without programming knowledge). A detailed example for a word processor was demonstrated in the second part. In this article, the PPTX format will be analyzed in a similar scenario. A simple injection with a hidden link to the active element is a demonstration of the undeclared capabilities of this type of file formats. It should be remembered that these files can be run or opened on absolutely any (including mobile) smart device. This article, like the previous ones, does not provide instructions for harming* firms or organizations**. The articles have exclusively educational functions and warnings for current specialists. 

This study provides a detailed analysis of penetration testing tools such as Metasploit, Core Impact, Immunity Canvas, and Security Forest. It compares their functionality, usability, and role in identifying vulnerabilities in information systems. The main focus is on the features of each tool, their strengths and weaknesses, and their areas of application. Penetration testing is an important element of a comprehensive approach to cybersecurity, allowing vulnerabilities to be identified at all stages of the information system lifecycle. The paper examines the stages of penetration testing, including information gathering, vulnerability identification, attack planning, and results analysis. Particular attention is paid to automated tools, which greatly simplify the testing process but require competent use. The article also discusses the ethical and legal aspects of penetration testing, emphasizing the need to comply with legislation and professional ethics. The article will be useful for information security specialists, as well as anyone interested in modern data protection methods. The paper emphasizes that the choice of tool depends on specific tasks and context, and that successful pentesting requires not only technical skills, but also a deep understanding of information protection processes.

The paper presents a demonstration of the analysis of a specific data set using machine learning methods to solve the research problem of finding a link between relative economic indicators (development indices) of the world’s countries according to the World Bank data for 2023 using machine learning methods and the use of structural equations.This approach was applied to the analysis of 205 countries using the weighted syndrome method and 180 countries using oneand twofactor confirmatory analysis. The advantage of using machine learning methods is the ability to use data with gaps, unlike regression models, which are the basis of factor analysis, when data gaps are not acceptable. To use the weighted syndromes method, eight main economic relative indicators of the countries’ development were used. The % GDP growth rate for 2023 was chosen as the grouping indicator. The quality of the model was assessed by the ROC AUK indicator. This indicator is 0.92, which indicates that the selected features really make it possible to divide the countries. Scattering diagrams showing a clear division of countries into two groups according to the analyzed indicators also illustrate the quality of recognition. The use of factor analysis made it possible to build two models (one-factor and two-factor) using not eight indicators, but only four, so that the models measured by the indices of comparative conformity and Tucker-Lewis were statistically significant. However, the loads of the one-factor model (λ) are extremely low (0.258, 0.131), which indicates a weak relationship between the observed variables and the latent factor. The factor practically does not explain the variance of the variables. These results indicate a poor correspondence of the model to the data, especially for the two—factor model, since the value of the TLI index is 0.474, which also cannot be a satisfactory result with an adequate threshold of >0.90–0.95). A meaningful economic interpretation of the latent factors obtained is not provided due to the poor correspondence of the results of factor analysis and the inability to compare machine learning and factor analysis methods with simultaneous analysis, since there was less data suitable for factor analysis than when using machine learning methods that demonstrated their adequacy.

Global changes in the global economy and trade place new demands on transport and logistics systems (TLS). Increasing their resilience to disruptions and structural changes is becoming a critical task amid the growing share of logistics costs in Russia’s GDP. Existing optimization methods often do not take into account the resilience of the network structure itself to destructive influences, which creates a gap in knowledge.

The purpose of the study is to develop and test methodological tools for assessing and improving the structural stability of the TL.

The approach integrates graph-theoretical modeling, multi-criteria optimization methods, and a new indicator, the percolation coefficient, which characterizes the network’s ability to deliver goods to all destinations. The multi-criteria optimization problem of finding paths and flows is formalized. Sustainability was assessed through the coefficient of influence of structural changes on the effectiveness of solutions. A large-scale computational experiment was conducted with the generation of more than 1 million graph structures. A mathematical model of the radar has been developed based on a matrix of initial conditions, and an efficiency coefficient has been proposed for comparing alternative options. A close correlation has been established between network bandwidth, percolation coefficient, and solution efficiency. The barrier values of the coefficient of influence have been determined, which make it possible to classify the system as stable or unstable to a specific type of structural failure. The principles of building sustainable radar stations are formulated, the key of which is the availability of alternative routes with efficiency close to optimal.

The results obtained lay the foundation for the creation of intelligent radar stations that are resistant to failures and load fluctuations.

The article is devoted to the analysis of approaches and solutions for the development of a single, scalable and highly loaded customer data warehouse within the bank’s digital ecosystem. In the context of digital transformation and the development of digital ecosystems, large organizations, including banks, are faced with the fact that the storage and processing of customer data is becoming a critical issue for ensuring company processes, personalization of services for customers, compliance with regulatory requirements and maintaining a competitive advantage. The purpose of the research is to determine the optimal set of architectural principles and technological solutions applicable to building a scalable, fault-tolerant and high-performance customer data storage system. The following methods were used to conduct the research: system analysis and synthesis to study the specifics of the subject area of the banking digital ecosystem, comparative analysis to evaluate data storage technologies based on criteria such as performance, scalability, data consistency, fault tolerance, and analysis of scientific literature and information sources to study the subject area, modern solutions and technologies in systems data storage. Based on the conducted research, optimal patterns and technologies for implementing a customer data storage platform have been selected and recommendations have been formulated for designing a single customer storage architecture capable of operating effectively in the dynamic IT landscape of the modern banking ecosystem. Based on the results of the study, it was concluded that other scientific papers and articles also confirm that the use of hybrid data warehouses using different technologies and solutions is most optimal for building highly loaded and scalable data storage and processing systems. The prospects of the research include further in-depth study of intelligent data management systems, data storage and processing used in machine learning and AI, Data Mesh and Data Fabric approaches and their application in the banking sector.

The article explores the applicability of fractal analysis in sociology and stock market analytics. The author considers the possibility of using the fractal principle to  study socio-economic indicators  and  forecast  stock market  activity. The purpose of the study is to demonstrate the potential of fractal methodology as an interdisciplinary analysis tool combining sociological observations, economic indicators and technical methods of stock market analysis. Main tasks: to analyze the theoretical foundations of the fractal approach; to identify the features of fractal structures in social and economic processes; to evaluate the effectiveness of fractal analysis in comparison with traditional methods of stock forecasting. The methodology includes: a review of the scientific literature on fractal analysis in finance and sociology; application of the concept of scalable invariance and fractional dimension; correlation analysis of dynamic series; modeling cyclic fractality on different horizons of expectations (micro, media and macrocycles). Results: The cyclical nature of social fractals and their ability to form sequences on the curve of median values of indicators have been confirmed. The scalability of fractal structures has been revealed, which makes it possible to expand the horizon of forecasting leading indicators. The correlation coefficient of dynamic series is calculated, demonstrating the stability of similarity regardless of the expectation horizon. It is shown that microcycles show a higher similarity to macrocycles, while the similarity to media cycles is less pronounced. Conclusions. Fractal analysis has advantages over traditional stock analysis methods (technical, fundamental) due to the numerical estimation of the probability of events, the volatility vector, the designation of trend reversal periods, and the scalability of invariants. The method is promising for macroeconomic forecasting, strategic planning and risk management, especially in combination with other indicators. However, fractal analysis is not universal: its effectiveness depends on the context and requires consideration of limitations (for example, the relativity of social time, the multifractality of data). The practical significance lies in the possibility of using fractal models to: refine stock forecasts; synthesize leading data; fill in the information gaps between regular reports (weekly, monthly).