BUSINESS INFORMATICS

Product matching in digital marketplaces: Multimodal model based on the transformer architecture

Artem Yu. Varnukhov — 2025-06-30

In this paper we analyze the problem of intelligent product matching in digital marketplaces for which one requires evaluation of similarity of various records that describe products but may differ in format, content or volume of multimodal data. The subject area of this scientific research represents an intersection of entity resolution (ER) problem solving methods: record matching and multimodal data analysis. It is of extreme relevance in a fast-growing platform economy with the e-commerce market expanding exponentially. The main purpose of this research is to develop and test an intelligent multimodal model based on transformer architecture to improve the accuracy and robustness of product matching in digital marketplaces. The authors developed a model integrating textual, visual and tabular attributes which enables us to identify similar products, find competitive offers, detect duplicates and perform product clustering and segmentation in a more effective manner. The proposed approach is based on the self-attention mechanism which enables contextual-semantic relations modeling of various-nature data. In order to extract the vector representation of text descriptions, language models are applied, in particular the Sentence-BERT architecture; for the graphical component Vision Transformer is used; and tabular data are processed using specialized learning mechanisms based on TabTransformer structured data. The experiment we carried out demonstrated that the developed multimodal model efficiently solves the task of product matching in digital marketplaces in an environment of significant variability of product items and data heterogeneity. Additionally, the results suggest that the model can be adapted successfully for application in other product categories. The results obtained confirm the efficiency and expediency to apply the multimodal approach for digital marketplace product matching implementation. This allows the e-commerce market participants to significantly improve the quality of inventory management, increase pricing efficiency and strengthen their competitive advantages.

Development of recommendation systems to improve the efficiency of regulated procurement in the electric power industry

Anna I. Denisova — 2025-06-30

This article considers ways to improve the efficiency of the regulated procurement market by implementing recommender systems into the existing procurement IT infrastructure. Using state, municipal and commercial procurement of electric power products as an example, the article considers promising classes of recommender systems for implementation, proposes a methodology for developing such services, and discloses algorithms for processing, configuring and interpreting data necessary for their operation. The difference between the author’s approach to creating services and previously published works is substantiated, testing and A/B testing are carried out, and an assessment of the effectiveness is presented. The results obtained have scientific novelty (the methodology of using neural networks in relation to the procurement industry has been substantiated) and practical significance (the customer’s time saved on searching for suppliers by up to 40%; the pool of potential suppliers has been expanded; supplier risks have been diversified by selecting relevant procedures from new areas and from new customers; suppliers have been provided with the opportunity to find up to 2–3 new customers for 1 recommendation mailing with a frequency of 1–2 times a week). We proposed to implement the developments in the practice of the operator of public procurement tenders. The authors see further development of recommendation services and solutions for the procurement industry in improving the analysis of semantic (text, logical) content of procurement documents, as well as the behavioral strategies of suppliers. The risks and limitations are associated with the high cost of maintaining a staff of developers-practitioners in neural networks, possible hallucinations of neural networks and their high sensitivity to errors and original data sets.

A customer avatar model based on Kolmogorov–Arnold networks

Fedor V. Krasnov — 2025-06-30

The increasing pace of development of e-commerce continues to present new challenges in terms of personalizing product search and recommendations. Monolithic search and recommendation systems have become cumbersome and are unable to effectively address the need for a deeper understanding of users on electronic trading platforms (ETPs) despite having access to comprehensive information about their interests and purchase histories. Collaborative filtering mechanisms which are widely used suffer from a lack of diversity in offerings and a reduced capacity to surprise users. Additionally, the low frequency of recommendation updates and the replacement of “personalized” with “similar to others” concepts contribute to these issues. We have approached the resolution of these issues by developing a shopping assistant named “Ellochka” that is individual for each user of ETP. The digital avatar model of the user continually searches for relevant products based on their history of interaction with ETP. We were guided by the principle of independence – avatar models do not share information with each other. When a new user joins, they are assigned a unique avatar model that evolves independently. Each avatar has its own language to generate search queries. The level of complexity of each avatar can vary depending on the intensity of its interaction with ETP. Continued interaction with the avatar allows for tracking of optimal purchase conditions, reminding users of expiration dates and the need for re-purchasing frequently purchased items. Isolating the avatar allows it to be retrained after each event, without significantly impacting the overall search and recommendation system. The use of neural network architecture-based and Kolmogorov–Arnold networks in the avatar-model has led to improvements in the main indicators of search and recommendation effectiveness, namely, novelty and diversity.

Mathematical model and intelligent system for analyzing the intensity of megaproject changes: the role of temporary change management hubs

Pavel A. Mikhnenko — 2025-06-30

Megaprojects represent large-scale investment programs with complex organizational structures, uniting a multitude of stakeholders whose interactions lead to the redistribution of power and the creation of temporary management centers. In conditions of unstable and uncertain external environments, such stakeholder behavior can result in the failure to achieve the set goals of the megaproject. An important scientific task is the development of mathematical models and methods for managing changes in megaprojects caused by the integrative actions of stakeholders under complex external conditions. The present study is aimed at creating a mathematical model and developing an information system for neural network analysis of the intensity of changes in megaprojects. Megaproject management is described using a vector-matrix model of a dynamic system with feedback based on the results of changes. To identify recurring patterns of negative events, the event-oriented analysis method was used. This allows for justifying new approaches to management aimed at reducing uncertainty and enhancing the effectiveness of megaproject implementation. Based on the proposed tools, a retrospective neural network analysis of the intensity of changes in the “Nord Stream 2” megaproject was conducted. Within the study, key groups of stakeholders were identified whose interactions significantly impacted the project’s implementation: Group 1 – Gazprom PJSC, European companies and the governments of Russia and Germany supporting the project; Group 2 – the governments of transit countries, the USA, environmental organizations and Baltic region countries opposing the project or expressing concern about its consequences. It was demonstrated that the integration of separate stakeholder groups contributes to the formation of temporary management centers with varying interests, leading to an increase in both positive and negative changes within the project. The outcome of the work was the development of an information system for analyzing the intensity of changes in megaprojects in the form of a prototype, which includes: a mathematical model for managing changes in megaprojects; a neural network analysis methodology based on the use of a large language model for processing textual information and generating quantitative assessments; as well as a software interface for uploading documents, automated data processing, and visualization of results. The primary neural network used was the large language model Qwen 2.5-Plus, which, while not specifically adapted for this task, had its parameters calibrated for analyzing the intensity of changes in megaprojects. The system prototype provides users with the ability to analyze stakeholder interactions, assess the intensity of changes and forecast potential risks based on historical data. A promising direction for further research involves applying the model we developed and neural network analysis methodology for comparative studies of various types of megaprojects.

Parallel implementation of the simplex method in matrix form using the PyTorch library for economics and management problems

Yuriy S. Ezrokh — 2025-06-30

The simplex method is widely used in economic planning and forecasting tasks. However, this method is used in real economic activity to find solutions to large-scale tasks, the speed of which is not a critical factor. This significantly limits the applied value of the simplex method in the economic sphere, since currently there is a certain tendency to move to more detailed economic models, which makes it urgent to accelerate calculations based on the simplex method. In these conditions, GPU (Graphical Processor Unit) computing accelerators become the most important means of accelerating calculations. The authors propose the implementation of the simplex method in matrix form for computing on GPUs using the PyTorch library. This allows you to switch to using the computing power of graphics processors in a simple and reliable way. A linear programming problem with 900 constraints is solved on a graphics accelerator 6–9 times faster than the solution on a conventional processor. This paper identifies groups of applied economic problems for which the proposed algorithms and methods can be relevant.

Ecosystem approach to strategic management on the example of agriculture

Vladimir I. Budzko — 2025-06-30

This paper considers the transformation of methods and models of strategic management based on the ecosystem approach within the framework of the formation of a unified digital platform for management. The ecosystem approach to the socio-economic development of society is gaining popularity as a result of global social requirements for environmental protection and a careful attitude to use of limited natural resources. Environmental problems in the Russian agricultural sector are increasing, in particular due to the process of formation of agro-industrial associations, mainly in the form of agricultural holdings. In this case, there is a problem of a systematic approach to using of technologies for the integration of all types of resources involved in production, taking into account the growing number and importance of environmental factors. Mathematical modeling is proposed as the main method of strategic management research. Unlike most of the existing models, which are often iconographic, the model proposed allows us to consider a larger number of factors. This makes it possible to assess various options of the modeling objects development using a simulation approach. As a result of the research, a mathematical model of strategic management of agroholdings for sustainable development was developed. It is shown that the development strategy should be implemented considering an appropriate automated management information system. This will lead to a radical change in the whole system of management and production. It will allow the enterprise to apply strategic goal-setting, focusing first of all on quality, controllability and other components of competitiveness. The mathematical model proposed provides justification of unified methods of long-term digitalization applicable for large agricultural associations, as well as for small and medium-sized farms, which will be able to cooperate with agricultural holdings relying on the principles of outsourcing.