Author: Andjela Djelic
Supervisor: Dominik Bork, Co-Supervisor: Julia Neidhardt
Abstract
In the Model-Driven Engineering (MDE) domain, Machine Learning (ML)-based recommender systems can assist developers during conceptual modeling by suggesting plausible modeling elements based on the current state of a model. However, the conceptual model datasets commonly used for such machine learning research, often mined from public software repositories, frequently suffer from significant quality issues, including duplicate models, trivial examples, and multilingual model element names.Beyond data quality, the multimodal nature of conceptual models, where information is conveyed through both semantic descriptions and structural relationships, poses an additional difficulty for the development of robust model completion tools. In particular, the transformation of Unified Modeling Language (UML) models into machine-readable representations often introduces a loss of information: approaches based on textual linearization disregard structural dependencies, whereas graph-based representations may lose the semantic richness of the model element names.This study introduces a comprehensive data cleansing pipeline specifically designed for UML model datasets, combining heuristics filtering for dummy models, similarity-based clone detection, and language filtering. In addition, a recommender system for UML class diagram completion is developed by fine-tuning a hybrid architecture that integrates Graph Neural Networks (GNNs) and Large Language Models (LLMs). This resulting model can recommend missing modeling elements, including class names, attributes, operations, and relationship types.The proposed approach is evaluated through reproducibility studies measuring the effect of cleansing on downstream ML tasks, as well as through experiments on UML class diagram completion tasks, comparing the performance of the fine-tuned model against a zero-shot baseline. The results demonstrate the potential of combining a UML-specific cleansing pipeline with a multimodal GNN-LLM auto-completion approach, pointing toward a more data-centric and structurally aware direction for machine learning in MDE.