我们提出了图注意力网络（GATs），这是一种新颖的神经网络架构，用于处理图结构数据。它利用掩蔽自注意力层来克服基于图卷积或其近似方法的先前方法的缺点。通过堆叠节点能够关注其邻域特征的层，我们能够（隐式地）为邻域中的不同节点指定不同的权重，而无需任何昂贵的矩阵运算（如求逆）或预先知道图结构。这样，我们同时解决了基于谱的图神经网络的几个关键挑战，并使我们的模型易于应用于归纳式以及直推式问题。我们的 GAT 模型在四个公认的直推式和归纳式图基准测试中取得了或匹配了最先进的结果：Cora、Citeseer 和 Pubmed 引文网络数据集，以及一个蛋白质-蛋白质相互作用数据集（其中测试图在训练期间未见）。
We present graph attention networks (GATs), novel neural network architectures that operate on graph-structured data, leveraging masked self-attentional layers to address the shortcomings of prior methods based on graph convolutions or their approximations. By stacking layers in which nodes are able to attend over their neighborhoods' features, we enable (implicitly) specifying different weights to different nodes in a neighborhood, without requiring any kind of costly matrix operation (such as inversion) or depending on knowing the graph structure upfront. In this way, we address several key challenges of spectral-based graph neural networks simultaneously, and make our model readily applicable to inductive as well as transductive problems. Our GAT models have achieved or matched state-of-the-art results across four established transductive and inductive graph benchmarks: the Cora, Citeseer and Pubmed citation network datasets, as well as a protein-protein interaction dataset (wherein test graphs remain unseen during training).
核心贡献 · Key contributions
提出使用掩码自注意力层的图注意力网络（GAT）处理图结构数据。 Proposes graph attention networks (GATs) using masked self-attentional layers for graph-structured data.
无需昂贵矩阵运算即可隐式地为邻域内节点分配不同权重。 Enables implicit assignment of different weights to nodes in a neighborhood without costly matrix operations.
在四个直推式和归纳式图基准上取得最先进结果。 Achieves state-of-the-art results on four transductive and inductive graph benchmarks.
引入多头注意力以稳定学习并提升模型容量。 Introduces multi-head attention to stabilize learning and improve model capacity.
展示了对完全未见测试图的归纳学习的适用性。 Demonstrates applicability to inductive learning with completely unseen test graphs.
通过学习到的注意力权重提供可解释性。 Provides interpretability through learned attention weights.
局限 · Limitations
稀疏矩阵操作限制了多图数据集的批处理能力。 Sparse matrix operations limit batching capabilities for multi-graph datasets.
对于稀疏图结构，GPU 性能可能不优于 CPU。 GPU performance may not surpass CPU for sparse graph structures.
感受野大小受网络深度限制，深层模型需跳跃连接。 Receptive field size is bounded by network depth, requiring skip connections for deeper models.
跨边并行化可能因邻域重叠导致冗余计算。 Parallelization across edges may cause redundant computation due to overlapping neighborhoods.
注意力机制的可解释性分析需领域知识，留待未来工作。 Attention mechanism analysis for interpretability requires domain knowledge and is left for future work.