The configuration process for feature-oriented product lines is well-researched for Boolean and numerical feature configurations. However, in several engineering fields, we encounter the challenge of finding the optimal graph structure to describe a product configuration. Optimising complex graph structures towards multiple objectives within numerous constraints requires a deep understanding of the graph configuration space and the product properties it represents. This study aims to leverage graph neural networks (GNNs) to predict product properties, thereby supporting the configuration process in product lines. In a controlled experiment, we compare a GNN-based approach to a recent state-of-the-art approach utilising graph embeddings. We evaluate these methods on both accuracy and learning efficiency. Our findings indicate that the GNN-based approach outperforms the embedding-based method in terms of accuracy. However, it requires a substantially larger volume of training data to achieve these results. Overall, this research demonstrates the applicability of an ML-supported framework for engineering product lines using graph configurations.
