Global Concept Explanations for Graphs by Contrastive Learning (2404.16532v1)

Abstract: Beyond improving trust and validating model fairness, xAI practices also have the potential to recover valuable scientific insights in application domains where little to no prior human intuition exists. To that end, we propose a method to extract global concept explanations from the predictions of graph neural networks to develop a deeper understanding of the tasks underlying structure-property relationships. We identify concept explanations as dense clusters in the self-explaining Megan models subgraph latent space. For each concept, we optimize a representative prototype graph and optionally use GPT-4 to provide hypotheses about why each structure has a certain effect on the prediction. We conduct computational experiments on synthetic and real-world graph property prediction tasks. For the synthetic tasks we find that our method correctly reproduces the structural rules by which they were created. For real-world molecular property regression and classification tasks, we find that our method rediscovers established rules of thumb. More specifically, our results for molecular mutagenicity prediction indicate more fine-grained resolution of structural details than existing explainability methods, consistent with previous results from chemistry literature. Overall, our results show promising capability to extract the underlying structure-property relationships for complex graph property prediction tasks.

• The paper introduces Megan2, which integrates contrastive learning and projection networks to generate global concept explanations for graph neural networks.
• It employs HDBSCAN clustering and genetic algorithms to extract and refine subgraph prototypes that capture key structural motifs.
• Computational validation on synthetic and real-world datasets demonstrates its effectiveness in uncovering structure-property relationships in chemistry and materials science.

Global Concept Explanations for Graphs by Contrastive Learning in Graph Neural Networks

Introduction

Graph neural networks (GNNs) have become pivotal in understanding and predicting properties from graph-structured data in domains like chemistry and material science. Existing methods predominantly focus on local explainability, which elucidates individual predictions. The novel approach discussed in this paper introduces a method termed Megan2, enhancing the earlier Megan model, to generate global concept explanations by identifying and clustering conceptually similar subgraph motifs across the graph's latent space.

Methodological Enhancements

The Megan2 model integrates a contrastive learning objective into the training of a multi-explanation graph attention network (Megan). The primary enhancements include:

• Projection Networks: These networks transform subgraph embeddings from each explanatory channel to promote independent evolution of channel representations, thereby enabling detailed encoding of subgraph similarities.
• Contrastive Learning Component: Through a modified training loss, this feature aims to align latent space embeddings with structural similarities between subgraphs. The method uses positive and negative sampling techniques to ensure that distances in the latent space reflect these similarities.
• Clustering for Concept Extraction: Applying the HDBSCAN clustering algorithm, Megan2 dynamically identifies dense clusters in the subgraph space as concepts. Each cluster essentially encases subgraphs exhibiting a shared structural pattern.
• Prototype and Hypothesis Generation: For each identified cluster, a prototype that represents a minimal subgraph encapsulating the concept is generated via a genetic algorithm. Additionally, when applicable, hypotheses about the structural effect on graph properties are generated using GPT-4, providing initial insights into potential causal relationships.

Computational Validation

The framework's capabilities were tested using both synthetic and real-world datasets. For synthetic datasets, Megan2 replicated known structure-property relationships accurately. In contrast, for molecular property prediction tasks, the model rediscovered established chemical thumb rules and unearthed more nuanced subgraph patterns contributing to properties like mutagenicity and solubility.

• Synthetic Datasets: The model was able to identify and reproduce the structural rules used to generate graph properties, confirming its ability to capture and generalize significant motifs.
• Real-World Datasets: The method’s application to datasets such as those for mutagenicity and water solubility predictions revealed its ability to condense complex pattern influences into understandable structural motifs. Notably, it offered more granular explanations in the field of mutagenicity by identifying multiple chemically relevant motifs that were consistent with literature but had been overlooked by previous explainability models.

Implications and Future Directions

This research paves the way for more transparent and interpretable AI models in graph structure analysis, particularly in fields where understanding the underlying structure-property relationships is crucial. The integration of global concept explanations can aid in debugging models, enhancing trust in AI-generated predictions, and potentially leading to new scientific insights where human intuition is limited.

Future work might extend these methodologies to more complex datasets and explore the impact of these concept explanations in assisting non-expert users. Moreover, it's imperative to refine the automatic hypothesis generation to mitigate the risk of propagating inaccuracies.

Conclusion

The proposed Megan2 framework marks a significant step towards explainable AI in graph neural networks by shifting the focus from local to global explanations. It crucially facilitates a deeper understanding of how GNNs make decisions and which graph substructures are most influential, thereby enhancing the model's transparency and usability for scientific discovery.