Poster Abstract: Graph Learning on Cyber-Physical Knowledge Graphs

Knowledge graphs (KGs) are machine-readable representations of cyber-physical systems (CPS) which can be incomplete due to the size and complexity of CPS. Knowledge graph completion (KGC) models can predict missing edges, but perform poorly and fail to generalize on CPS KGs due to the high heterogeneity and small size of those graphs. In this work, we propose an ontology-informed heterogeneous Graph Neural Network (GNN) architecture that integrates hierarchical parent-class layers to enhance generalization in CPS KGs. Our approach outperforms traditional heterogeneous GNNs and Node2Vec-based methods in edge prediction tasks, offering a promising solution for CPS KGs.