Path reasoning graph neural networks achieve strong results on knowledge graph completion tasks like link prediction, but their predictions remain difficult to interpret. Most existing GNN explainability methods are designed for node-level or graph-level tasks, leaving a gap for link prediction in knowledge graphs.
RAW-Explainer fills that gap by generating connected, concise subgraph explanations for individual predictions. The method leverages heterogeneous information to identify factual explanations and uses a random walk objective to encourage connectivity, with a parameterized model that produces explanations collectively and efficiently.
A key contribution is a robust evaluator that generalizes to the subgraph distribution, overcoming the distribution shift that occurs when explanatory subgraphs are orders of magnitude smaller than the full graph. Experiments on standard knowledge graph completion benchmarks show that RAW-Explainer balances explanation quality with computational efficiency.
The paper, accepted at the Learning on Graphs Conference 2026, makes code and data publicly available.