Explaining Graph Neural Networks with Large Language Models: A Counterfactual Perspective for Molecular Property Prediction (2410.15165v1)

Abstract: In recent years, Graph Neural Networks (GNNs) have become successful in molecular property prediction tasks such as toxicity analysis. However, due to the black-box nature of GNNs, their outputs can be concerning in high-stakes decision-making scenarios, e.g., drug discovery. Facing such an issue, Graph Counterfactual Explanation (GCE) has emerged as a promising approach to improve GNN transparency. However, current GCE methods usually fail to take domain-specific knowledge into consideration, which can result in outputs that are not easily comprehensible by humans. To address this challenge, we propose a novel GCE method, LLM-GCE, to unleash the power of LLMs in explaining GNNs for molecular property prediction. Specifically, we utilize an autoencoder to generate the counterfactual graph topology from a set of counterfactual text pairs (CTPs) based on an input graph. Meanwhile, we also incorporate a CTP dynamic feedback module to mitigate LLM hallucination, which provides intermediate feedback derived from the generated counterfactuals as an attempt to give more faithful guidance. Extensive experiments demonstrate the superior performance of LLM-GCE. Our code is released on https://github.com/YinhanHe123/new\_LLM4GNNExplanation.

• The paper introduces LLM-GCE, a novel framework that combines LLMs with GNNs to generate coherent counterfactual explanations in molecular property prediction.
• The methodology employs a counterfactual autoencoder and a dynamic feedback module to mitigate LLM hallucinations and incorporate domain-specific knowledge.
• Experimental results on five real-world datasets demonstrate that LLM-GCE significantly improves validity and proximity scores in generating feasible counterfactuals.

LLMs for Guiding Graph Neural Network Explanations: A Counterfactual Approach

The paper presents an investigation into leveraging LLMs to enhance the explainability of Graph Neural Networks (GNNs) in molecular property prediction tasks. The need for explainability in GNNs is acute in high-stakes fields such as drug discovery, where understanding model predictions can significantly impact decision-making processes.

Methodological Innovations

The authors introduce LLM-GCE, a novel framework combining the text generation strengths of LLMs with GNNs for graph counterfactual explanation (GCE). This approach aims to address two prevalent issues in current GCE models: the generation of incomprehensible counterfactuals and the disregard of domain-specific knowledge, resulting in less reliable predictions.

Key Components of LLM-GCE:

1. Counterfactual Autoencoder (CA):
   • Utilizes an autoencoder supported by a BERT-based text encoder to map generated counterfactual text pairs (CTPs) into a latent space.
   • Integrates a graph decoder to generate modifications on input graphs aligned with CTPs, resulting in counterfactual graphs.
2. Dynamic Feedback Module:
   • Mitigates potential LLM hallucinations by providing iterative feedback on generated counterfactuals. This adjustment is informed by the GNN's prediction outcomes on the modifications applied, facilitating more accurate and realistic counterfactuals.

Empirical Evaluation

The effectiveness of LLM-GCE is thoroughly validated through experiments on five real-world datasets pertinent to molecular property prediction. The results consistently demonstrate that LLM-GCE achieves higher validity and proximity scores in generating feasible counterfactuals compared to established GCE techniques.

Implications for Future Research

The introduction of LLMs into the field of GNN explanations is a promising venture, offering a new dimension to interpretability in machine learning models. This paper's methodologies can potentially extend beyond molecular graphs to include other data types where GNNs are applicable, such as social networks or recommendation systems.

Challenges and Considerations

Noteworthy challenges include the computational intensity associated with LLMs and the inherent risk of bias present in pre-trained LLMs, which may transfer into the GCE outputs. Additionally, scalability remains a concern, particularly when addressing complex, large-scale graph data.

Conclusion

LLM-GCE indicates a significant advancement in GNN explanation techniques, facilitating the production of more intuitive and domain-consistent counterfactuals. The framework not only enhances transparency in machine learning models but also opens up discussions for further leveraging LLM capabilities in diverse analytical applications.

This paper thus sets the groundwork for incorporating complex textual narratives into GNN models, broadening the horizons for more transparent AI systems.