Deep learning architectures for learning interdependencies in knowledge graphs

Integrating structured knowledge with neural learning systems is an important challenge in Artificial Intelligence. Knowledge Graphs (KGs) represent real-world knowledge as networks of entities connected by typed relations, and they underpin applications ranging from search engines to drug discovery. Current deep learning approaches for KGs, however, struggle to learn representations that capture the temporal, conditional and semantic interdependencies found in real-world Knowledge Graphs. This thesis investigates how deep learning systems can effectively learn representations of interdependent triples in Knowledge Graphs, and makes four contributions. First, a reproduction study of Relational Graph Convolutional Networks introduces two parameter-efficient variants, one of which achieves competitive node-classification accuracy with only 8% of the original model's parameters. Second, A-NeSI is a neurosymbolic inference framework that replaces #P-hard exact Weighted Model Counting with approximate inference that scales polynomially in the number of ground atoms, with a symbolic pruner guaranteeing constraint satisfaction by construction. Third, IntelliGraphs is a benchmark suite showing that traditional KG embedding models almost never generate semantically valid subgraphs, with validity below 1% on every dataset tested. Fourth, ARK and SAIL are autoregressive sequence models that reach between 89.22% and 100% semantic validity on all five IntelliGraphs datasets, with single-layer recurrent architectures matching the validity of deeper transformer variants. Together, these findings show that capturing interdependencies in Knowledge Graphs depends less on architectural complexity than on whether a model's inductive bias matches the structure of the problem.