Generated Distributed Code from Single Node Code (Poster Practice)

Migrating monolithic systems to microservices is a challenging task due to tightly coupled components and unclear service boundaries. Existing decomposition approaches often rely on either structural dependencies or semantic representations, limiting their ability to capture the full system context. This work proposes an automated framework that integrates structural and semantic signals using a graph-based learning approach. The system models the monolith as a heterogeneous graph and employs an edge-aware graph autoencoder to learn meaningful embeddings of program entities. Triplet loss is incorporated to preserve semantic similarity, while K-Means clustering is used to identify microservice boundaries. The framework is evaluated on benchmark datasets using standard metrics such as Structural Modularity (SM), Inter-Service Coupling (ICP), Normalized Entity Distribution (NED), and Interface Number (IFN). Results demonstrate improved decomposition quality and more balanced service partitions compared to existing approaches.