แนวคิดหลัก
Large language models (LLMs) can be effectively used to guide the generation of more realistic and interpretable counterfactual explanations for graph neural networks (GNNs) in molecular property prediction, improving the transparency and trustworthiness of GNNs in this domain.
สถิติ
The graph label distribution in the Tox21 dataset is heavily skewed, with more than 95% of the labels being 0.
For the Tox21 dataset, 600 zero-labeled graphs were randomly selected to address the skewed label distribution.
The study used five real-world datasets: AIDS, Mutagenicity, BBBP, ClinTox, and Tox21.
Molecules with only one atom or more than 100 atoms were removed during data preprocessing.
The GNN classifier achieved accuracies above 70% on all datasets, with the highest accuracy being 99.4% on the AIDS dataset.
LLM-GCE achieved the highest validity among almost all baselines across all datasets when considering chemical feasibility.
LLM-GCE consistently produced counterfactuals with lower proximity compared to baseline methods, indicating greater similarity to the original molecules.
คำพูด
"To handle the above limitations, large language models (LLMs) [...] are ideal for addressing these limitations due to their ability to (i) generate comprehensible natural language texts, (ii) make the counterfactual optimization process human-interpretable, and (iii) leverage inherent domain knowledge from extensive pretraining to produce realistic counterfactuals."
"LLM-GCE unlocks LLM’s strong reasoning ability in GCE by addressing hallucinations and graph structure inference limitations."
"Extensive experiments demonstrate the superior performance of LLM-GCE."