Explainability in Graph Neural Networks

Graph Neural Networks (GNNs) have emerged as a powerful class of models for learning from graph-structured data, capturing complex relational patterns across nodes and edges. However, their inherent complexity often renders them opaque, hampering trust in critical domains such as drug discovery, social network analysis and financial risk assessment. Explainability in GNNs seeks to make the decision process transparent by identifying which nodes, edges or substructures drive a given prediction. Approaches range from post-hoc techniques, which analyse fully trained models to extract salient subgraphs, to intrinsically interpretable architectures that incorporate explanation mechanisms during training. Evaluation frameworks and synthetic benchmark generators have been introduced to assess the faithfulness and robustness of explanations. At the same time, methods that quantify uncertainty in explanations are gaining traction, ensuring that users understand the confidence and potential variability of the insights provided. Together, these advances advance the global understanding of GNN behaviour, support compliance with regulatory requirements and foster broader adoption in applications demanding both high performance and clear rationale.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Higher-Order Explanations of Graph Neural Networks via Relevant Walks introduces a nested attribution scheme that extends layer-wise relevance propagation to identify joint contributions of groups of edges. By extracting relevant walks—sequences of edges that most influence a prediction—this method reveals higher-order interactions in domains such as quantum chemistry and sentiment analysis, offering a more nuanced interpretation than single-edge attributions.

Evaluating explainability for graph neural networks presents ShapeGGen, a synthetic graph generator that produces benchmark datasets with ground-truth explanations. Coupled with GraphXAI, a library offering data loaders, model implementations and evaluation metrics, this work enables systematic comparison of explainers under varied graph topologies and homophily settings, promoting standardisation in the assessment of GNN interpretability.

L2xGnn: learning to explain graph neural networks proposes a faithful-by-design framework that jointly trains a selector to choose sparse, connected subgraphs (motifs) and a GNN that makes predictions using only those motifs. By enforcing that explanations drive the model’s message-passing operations, the approach ensures consistency between the explanation and the underlying decision, achieving competitive accuracy while reducing explanation complexity.

Explainability in Graph Neural Networks publication trend

The graph below shows the total number of articles in explainability in graph neural networks across all publications each year (not limited to Nature Index journals).

Technical terms

Graph Neural Network: A neural architecture that processes graph-structured data by aggregating and transforming information across nodes and edges.
Post-hoc explanation: A technique applied after model training to infer which inputs or substructures most influence a prediction.
Message passing: The iterative mechanism by which nodes in a graph exchange and update feature information along edges.
Subgraph motif: A connected subset of nodes and edges identified as a basis for explaining model decisions.
Ground-truth explanation: A reference annotation or synthetic pattern considered to accurately represent the true drivers of a model’s prediction.

References

1. Higher-Order Explanations of Graph Neural Networks via Relevant Walks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022).
2. Evaluating explainability for graph neural networks. Scientific Data (2023).
3. L2XGNN: learning to explain graph neural networks. Machine Learning (2024).
4. Quantifying uncertainty in graph neural network explanations. Frontiers in Big Data (2024).