Global Concept Explanations for Graphs by Contrastive Learning

The method successfully identifies meaningful structural patterns: When tested on synthetic tasks, the approach correctly reproduced the structural rules used to create the data, and on real-world molecular tasks, it rediscovered established chemistry "rules of thumb," suggesting the explanations capture genuine domain knowledge rather than spurious correlations.

Graph neural network explanations can reveal scientific insights: Beyond validating that AI models work fairly, this explainability approach recovered valuable scientific understanding in domains where human intuition is limited, demonstrating xAI's potential as a discovery tool rather than just a trust-building mechanism.

The method provides finer-grained chemical insights than existing approaches: For molecular mutagenicity prediction specifically, the technique achieved "more fine-grained resolution of structural details than existing explainability methods," aligning with established chemistry literature and suggesting it could help scientists understand molecular properties more deeply.

Concept explanations work by identifying patterns in AI model internals: The approach extracts global explanations by finding dense clusters in the neural network's latent space and creating representative prototype graphs, making the implicit patterns the model learned explicit and interpretable.

AI explanations can be enhanced with language models: The method optionally uses GPT-4 to generate hypotheses about why specific molecular structures affect predictions, combining neural network interpretability with natural language to make findings more accessible to domain experts.