A METHOD OF AUTOMATED FORMATION OF CLUSTERS OF THE SOFTWARE SYSTEM BASED ON STRUCTURAL-SEMANTIC ANALYSIS OF COMPONENTS USING NEURAL NETWORKS

Abstract

The article examines the problem of transformation of monolithic architectures of software systems into microservice ones, which is one of the key tasks of modern software engineering. Monolithic systems eventually acquire excessive complexity and a high level of internal connectivity, making them difficult to scale, maintain and modify. Microservice architecture, oriented towards independent services with clear interfaces, largely solves these problems, however, the non-trivial task of correct decomposition arises, which does not have a single universal solution.The paper proposes a decomposition method that combines structural and semantic analysis of source code with training of neural networks for graphs aimed at automated formation of the cluster structure of the system. A feature of the approach is the multi-level construction of the dependency graph, which takes into account different types of relationships (calls of methods, imitation, use of resources), as well as the introduction of node weights calculated on the basis of centrality indicators. Prior to the clustering step, a stepwise structural graph optimization is applied, involving the detection of strongly coupled components, cycles, bridges and articulation points, and the reduction of the influence of redundant and noise links.This makes it possible to increase the accuracy and reliability of further grouping of components.In addition, an adaptive selection of the clustering algorithm based on graph metrics (density, number of nodes, centrality, resource sharing coefficient) is implemented, which allows you to match the method with the current characteristics of the structure. Graph neural network training is carried out on feature vectors that integrate topological and semantic characteristics of components, which ensures the detection of latent regularities in complex dependencies.Comparative analysis with current decomposition methods (CoGCN, DEEPLY, HyDEC, GDC-DVF, Mo2oM) confirmed the effectiveness of the proposed approach. The results showed an excess of existing solutions according to the structural modularity indicator (SM), which reflects higher cohesion and clearer boundaries of the formed services, while maintaining a relatively low level of interservice dependencies. This shows the scientific novelty and practical significance of the developed method, which can be used both in the research environment and in real industrial systems.