
摘要
基于消息传递(Message Passing, MP)范式的图神经网络(Graph Neural Networks, GNNs)通常仅在1跳邻居之间交换信息,以逐层构建节点表示。原则上,这类网络难以捕捉可能对特定图学习任务至关重要或不可或缺的长程相互作用(Long-Range Interactions, LRI)。近年来,基于Transformer的图方法受到越来越多关注,这类方法能够超越原始图结构的稀疏连接,考虑节点间的全连接关系,从而支持对LRI的建模。然而,在现有多个图基准测试中,仅依赖1跳消息传递的MP-GNNs 若结合位置特征表示等其他创新技术,往往表现更优,这在一定程度上限制了类Transformer架构的实用价值认知及其在性能排名中的地位。为此,我们提出了长程图基准测试集(Long Range Graph Benchmark, LRGB),包含5个图学习数据集:PascalVOC-SP、COCO-SP、PCQM-Contact、Peptides-func 与 Peptides-struct。这些数据集在任务表现上明显依赖于对长程相互作用的推理能力,因此可作为评估LRI建模能力的有效基准。我们在LRGB上对基础GNN模型与图Transformer网络进行了系统性对比实验,结果表明,能够有效捕捉长程依赖关系的模型在这些任务上显著优于传统方法。因此,LRGB数据集为评估和探索旨在建模长程相互作用的MP-GNN与图Transformer架构提供了理想的基准平台。
代码仓库
vijaydwivedi75/lrgb
官方
pytorch
GitHub 中提及
zml72062/dr-fwl-2
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| graph-classification-on-peptides-func | GCN | AP: 0.5930±0.0023 |
| graph-classification-on-peptides-func | GINE | AP: 0.5498±0.0079 |
| graph-classification-on-peptides-func | GatedGCN | AP: 0.5864±0.0077 |
| graph-classification-on-peptides-func | Transformer+LapPE | AP: 0.6326±0.0126 |
| graph-classification-on-peptides-func | GatedGCN+RWSE | AP: 0.6069±0.0035 |
| graph-classification-on-peptides-func | SAN+LapPE | AP: 0.6384±0.0121 |
| graph-classification-on-peptides-func | SAN+RWSE | AP: 0.6439±0.0075 |
| graph-regression-on-peptides-struct | Transformer+LapPE | MAE: 0.2529±0.0016 |
| graph-regression-on-peptides-struct | SAN+LapPE | MAE: 0.2683±0.0043 |
| graph-regression-on-peptides-struct | GCN | MAE: 0.3496±0.0013 |
| graph-regression-on-peptides-struct | GINE | MAE: 0.3547±0.0045 |
| graph-regression-on-peptides-struct | GatedGCN | MAE: 0.3420±0.0013 |
| graph-regression-on-peptides-struct | GatedGCN+RWSE | MAE: 0.3357±0.0006 |
| graph-regression-on-peptides-struct | SAN+RWSE | MAE: 0.2545±0.0012 |
| link-prediction-on-pcqm-contact | GCN | Hits@1: 0.1321±0.0007 Hits@10: 0.8256±0.0006 Hits@3: 0.3791±0.0004 MRR: 0.3234±0.0006 |
| link-prediction-on-pcqm-contact | SAN+LapPE | Hits@1: 0.1355±0.0017 Hits@10: 0.8478±0.0044 Hits@3: 0.4004±0.0021 MRR: 0.3350±0.0003 |
| link-prediction-on-pcqm-contact | Transformer+LapPE | Hits@1: 0.1221±0.0011 Hits@10: 0.8517±0.0039 Hits@3: 0.3679±0.0033 MRR: 0.3174±0.0020 |
| link-prediction-on-pcqm-contact | SAN+RWSE | Hits@1: 0.1312±0.0016 Hits@10: 0.8550±0.0024 Hits@3: 0.4030±0.0008 MRR: 0.3341±0.0006 |
| link-prediction-on-pcqm-contact | GatedGCN+RWSE | Hits@1: 0.1288±0.0013 Hits@10: 0.8517±0.0005 Hits@3: 0.3808±0.0006 MRR: 0.3242±0.0008 |
| link-prediction-on-pcqm-contact | GatedGCN | Hits@1: 0.1279±0.0018 Hits@10: 0.8433±0.0011 Hits@3: 0.3783±0.0004 MRR: 0.3218±0.0011 |
| link-prediction-on-pcqm-contact | GINE | Hits@1: 0.1337±0.0013 Hits@10: 0.8147±0.0062 Hits@3: 0.3642±0.0043 MRR: 0.3180±0.0027 |
| node-classification-on-coco-sp | GatedGCN | macro F1: 0.2641±0.0045 |
| node-classification-on-coco-sp | GatedGCN+LapPE | macro F1: 0.2574±0.0034 |
| node-classification-on-coco-sp | GCN | macro F1: 0.0841±0.0010 |
| node-classification-on-coco-sp | GINE | macro F1: 0.1339±0.0044 |
| node-classification-on-coco-sp | SAN+RWSE | macro F1: 0.2434±0.0156 |
| node-classification-on-coco-sp | Transformer+LapPE | macro F1: 0.2618±0.0031 |
| node-classification-on-coco-sp | SAN+LapPE | macro F1: 0.2592±0.0158 |
| node-classification-on-pascalvoc-sp-1 | GatedGCN | macro F1: 0.2873±0.0219 |
| node-classification-on-pascalvoc-sp-1 | SAN+RWSE | macro F1: 0.3216±0.0027 |
| node-classification-on-pascalvoc-sp-1 | GCN | macro F1: 0.1268±0.0060 |
| node-classification-on-pascalvoc-sp-1 | Transformer+LapPE | macro F1: 0.2694±0.0098 |
| node-classification-on-pascalvoc-sp-1 | GINE | macro F1: 0.1265±0.0076 |
| node-classification-on-pascalvoc-sp-1 | SAN+LapPE | macro F1: 0.3230±0.0039 |
| node-classification-on-pascalvoc-sp-1 | GatedGCN+LapPE | macro F1: 0.2860±0.0085 |