4 个月前

一种端到端的基于注意力机制的图学习方法

一种端到端的基于注意力机制的图学习方法

摘要

近期,基于变换器的架构在图学习领域得到了迅速发展,这主要是由于注意力机制作为一种有效的学习方法以及希望取代消息传递方案中手工设计的操作者。然而,人们对其经验有效性、可扩展性和预处理步骤的复杂性提出了担忧,尤其是在与通常在广泛基准测试中表现相当的更简单的图神经网络相比时。为了解决这些不足,我们将图视为边的集合,并提出了一种纯基于注意力的方法,该方法由编码器和注意力池化机制组成。编码器纵向交替使用掩码自注意力模块和普通自注意力模块来学习边的有效表示,同时允许处理输入图中的可能错误指定问题。尽管该方法简单,但在超过70个节点级和图级任务上超越了精心调校的消息传递基线模型和最近提出的基于变换器的方法,包括具有挑战性的长程基准测试。此外,我们在不同任务中展示了最先进的性能,从分子图到视觉图再到异质节点分类。该方法在迁移学习设置中也优于图神经网络和变换器,并且其可扩展性远胜于性能水平或表达能力相似的其他替代方案。

基准测试

基准方法指标
graph-classification-on-cifar10-100kESA (Edge set attention, no positional encodings)
Accuracy (%): 75.413±0.248
graph-classification-on-ddESA (Edge set attention, no positional encodings)
Accuracy: 83.529±1.743
graph-classification-on-enzymesESA (Edge set attention, no positional encodings)
Accuracy: 79.423±1.658
graph-classification-on-imdb-bESA (Edge set attention, no positional encodings)
Accuracy: 86.250±0.957
graph-classification-on-malnet-tinyESA (Edge set attention, no positional encodings)
Accuracy: 94.800±0.424
MCC: 0.935±0.005
graph-classification-on-mnistESA (Edge set attention, no positional encodings)
Accuracy: 98.753±0.041
graph-classification-on-mnistESA (Edge set attention, no positional encodings, tuned)
Accuracy: 98.917±0.020
graph-classification-on-nci1ESA (Edge set attention, no positional encodings)
Accuracy: 87.835±0.644
graph-classification-on-nci109ESA (Edge set attention, no positional encodings)
Accuracy: 84.976±0.551
graph-classification-on-peptides-funcESA (Edge set attention, no positional encodings, not tuned)
AP: 0.6863±0.0044
graph-classification-on-peptides-funcESA (Edge set attention, no positional encodings, tuned)
AP: 0.7071±0.0015
graph-classification-on-peptides-funcESA + RWSE (Edge set attention, Random Walk Structural Encoding, tuned)
AP: 0.7357±0.0036
graph-classification-on-peptides-funcESA + RWSE (Edge set attention, Random Walk Structural Encoding, + validation set)
AP: 0.7479
graph-classification-on-proteinsESA (Edge set attention, no positional encodings)
Accuracy: 82.679±0.799
graph-regression-on-esr2ESA (Edge set attention, no positional encodings)
R2: 0.697±0.000
RMSE: 0.486±0.697
graph-regression-on-f2ESA (Edge set attention, no positional encodings)
R2: 0.891±0.000
RMSE: 0.335±0.891
graph-regression-on-kitESA (Edge set attention, no positional encodings)
R2: 0.841±0.000
RMSE: 0.433±0.841
graph-regression-on-lipophilicityESA (Edge set attention, no positional encodings)
R2: 0.809±0.008
RMSE: 0.552±0.012
graph-regression-on-parp1ESA (Edge set attention, no positional encodings)
R2: 0.925±0.000
RMSE: 0.343±0.925
graph-regression-on-pcqm4mv2-lscESA (Edge set attention, no positional encodings)
Test MAE: N/A
Validation MAE: 0.0235
graph-regression-on-peptides-structESA + RWSE (Edge set attention, Random Walk Structural Encoding, tuned)
MAE: 0.2393±0.0004
graph-regression-on-peptides-structESA (Edge set attention, no positional encodings, not tuned)
MAE: 0.2453±0.0003
graph-regression-on-pgrESA (Edge set attention, no positional encodings)
R2: 0.725±0.000
RMSE: 0.507±0.725
graph-regression-on-zincESA + rings + NodeRWSE + EdgeRWSE
MAE: 0.051
graph-regression-on-zinc-500kESA + rings + NodeRWSE + EdgeRWSE
MAE: 0.051
graph-regression-on-zinc-fullESA + rings + NodeRWSE + EdgeRWSE
Test MAE: 0.0109±0.0002
graph-regression-on-zinc-fullESA + RWSE (Edge set attention, Random Walk Structural Encoding, tuned)
Test MAE: 0.0154±0.0001
graph-regression-on-zinc-fullESA + RWSE (Edge set attention, Random Walk Structural Encoding)
Test MAE: 0.017±0.001
graph-regression-on-zinc-fullESA + RWSE + CY2C (Edge set attention, Random Walk Structural Encoding, clique adjacency, tuned)
Test MAE: 0.0122±0.0004
graph-regression-on-zinc-fullESA (Edge set attention, no positional encodings)
Test MAE: 0.027±0.001
molecular-property-prediction-on-esolESA (Edge set attention, no positional encodings)
R2: 0.944±0.002
RMSE: 0.485±0.009
molecular-property-prediction-on-freesolvESA (Edge set attention, no positional encodings)
R2: 0.977±0.001
RMSE: 0.595±0.013

用 AI 构建 AI

从想法到上线——通过免费 AI 协同编程、开箱即用的环境和市场最优价格的 GPU 加速您的 AI 开发

AI 协同编程
即用型 GPU
最优价格
立即开始

Hyper Newsletters

订阅我们的最新资讯
我们会在北京时间 每周一的上午九点 向您的邮箱投递本周内的最新更新
邮件发送服务由 MailChimp 提供
一种端到端的基于注意力机制的图学习方法 | 论文 | HyperAI超神经