
摘要
图神经网络(Graph Neural Networks, GNNs)已被证明在处理图结构数据的各类预测任务中具有优异性能。近期关于其表达能力的研究主要聚焦于同构性判定任务以及可数特征空间。本文将这一理论框架拓展至包含连续特征的情形——这类特征在现实世界的输入数据中频繁出现,同时也广泛存在于GNN的隐藏层中,并在此背景下揭示了在多聚合函数需求上的必要性。基于此,我们提出了一种新型架构——主邻域聚合(Principal Neighbourhood Aggregation, PNA),该架构结合了多种聚合函数与度数缩放机制(degree-scalers),后者是对求和聚合器的广义化。最后,我们通过一个新颖的基准测试体系,系统评估了不同模型在捕捉与利用图结构方面的性能,该基准涵盖来自经典图论的多个任务,以及来自真实世界领域的现有基准。实验结果一致表明,所提出的PNA模型展现出显著优势。本研究旨在推动图神经网络领域的研究方向,朝向新型聚合方法的探索,我们认为这些方法对于构建强大且鲁棒的图神经网络模型至关重要。
代码仓库
dmlc/dgl
pytorch
GitHub 中提及
Saro00/DGN
pytorch
asarigun/GraphMixerNetworks
pytorch
GitHub 中提及
rusty1s/pytorch_geometric
pytorch
GitHub 中提及
changminwu/expandergnn
pytorch
GitHub 中提及
cvignac/SMP
pytorch
GitHub 中提及
lukecavabarrett/pna
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| graph-classification-on-cifar10-100k | PNA | Accuracy (%): 70.47 |
| graph-classification-on-dd | PNA | Accuracy: 78.992±4.407 |
| graph-classification-on-enzymes | PNA | Accuracy: 73.021±2.512 |
| graph-classification-on-imdb-b | PNA | Accuracy: 78.000±3.808 |
| graph-classification-on-nci1 | PNA | Accuracy: 84.964±1.391 |
| graph-classification-on-nci109 | PNA | Accuracy: 83.382±1.045 |
| graph-classification-on-proteins | PNA | Accuracy: 77.679±3.281 |
| graph-property-prediction-on-ogbg-molhiv | PNA | Ext. data: No Number of params: 326081 Test ROC-AUC: 0.7905 ± 0.0132 Validation ROC-AUC: 0.8519 ± 0.0099 |
| graph-property-prediction-on-ogbg-molpcba | PNA | Ext. data: No Number of params: 6550839 Test AP: 0.2838 ± 0.0035 Validation AP: 0.2926 ± 0.0026 |
| graph-regression-on-esr2 | PNA | R2: 0.696±0.000 RMSE: 0.486±0.696 |
| graph-regression-on-f2 | PNA | R2: 0.891±0.000 RMSE: 0.336±0.891 |
| graph-regression-on-kit | PNA | R2: 0.843±0.000 RMSE: 0.430±0.843 |
| graph-regression-on-lipophilicity | PNA | R2: 0.830±0.007 RMSE: 0.520±0.011 |
| graph-regression-on-parp1 | PNA | R2: 0.924±0.000 RMSE: 0.346±0.924 |
| graph-regression-on-pgr | PNA | R2: 0.717±0.000 RMSE: 0.514±0.717 |
| graph-regression-on-zinc | PNA | MAE: 0.142 |
| graph-regression-on-zinc-full | PNA | Test MAE: 0.057±0.007 |
| molecular-property-prediction-on-esol | PNA | R2: 0.942±0.006 RMSE: 0.493±0.026 |
| molecular-property-prediction-on-freesolv | PNA | R2: 0.951±0.009 RMSE: 0.870±0.081 |
| node-classification-on-pattern-100k | PNA | Accuracy (%): 86.567 |