Leveraging Machine Learning and Transformers to Identify Domain-Specific Services Decomposition in Legacy Systems

Service-oriented architecture, one of the popular software architectures that have become very popular in recent years, has scalability, isolation and flexibility as it consists of smaller and independent domain-specific services compared to monolithic systems. For this reason, the transition from monolithic monolithic systems to service-oriented architectures is becoming widespread for large-scale applications with millions of users to have an easily manageable, scalable and flexible structure. In this study, the effectiveness of various machine learning models and different types of tokenization methods were evaluated by analyzing static source code to decompose monolithic legacy systems into domain-specific services. Standard machine learning algorithms and transformer-based tokenizers were applied to the FXML-POS legacy system and model performance were evaluated using precision, recall, accuracy, and F1 scores. Experimental results indicate that all transformer models achieve strong performance with an F1 score of 91.9% using Random Forest and Logistic Regression classifiers. Furthermore, it has been observed in the experimental results that the Word2Vec vectorization method outperforms TF-IDF in most scenarios and a maximum F1 score of 97.2% is achieved using the Random Forest Classifier. These results underscore the utility of advanced embedding techniques and classifiers in the accurate identification of domain-specific service components.