Tag
该论文提出 Graph-Dependent Pooling (GDPool)，一种动态决定每个图池化粒度的层级图神经网络池化方法，通过自注意力选择锚点并在金属有机框架和分子图基准上取得最佳性能。
The paper investigates why Graph Transformers rarely form attention sinks compared to LLMs, showing that the attention scoring function — not graph structure — is the main cause: replacing GRIT's gated logit with dot-product attention raises its sink rate from 0.09 to 0.68.
This paper identifies a fundamental issue in top-k explanations for graph neural networks: automorphisms in input graphs cause non-unique explanations, as the model cannot distinguish symmetric elements. The authors provide a criterion to detect such arbitrariness and verify it using automated reasoning in Lean 4, showing the problem is widespread in molecular datasets.
This paper proposes ChemHyperMag, a physics-informed magnetic hypergraph learning method for multitask ADMET prediction that uses functional group hypergraphs and a Hermitian magnetic Laplacian to capture asymmetric interactions and directional signals, improving prediction accuracy with fewer labeled samples.