Graph Neural Networks~(GNNs) are effective tools for graph representation
learning. Most GNNs rely on a recursive neighborhood aggregation scheme, named
message passing, thereby their theoretical expressive power is limited to the
first order Weisfeiler-Lehman test (1-WL). Motivated by the success of
retrieval-based models and off-the-shelf high-performance retrieval systems, we
propose a non-parametric and model-agnostic scheme called GraphRetrieval to
boost existing GNN models. In GraphRetrieval, similar training graphs
associated with their ground-truth labels are retrieved as an enhancement to be
jointly utilized with the input graph representation to complete various graph
property predictive tasks. In particular, to effectively "absorb" useful
information from retrieved graphs and "ignore" possible noise, we introduce an
adapter based on self-attention to explicitly learn the interaction between an
input graph and its retrieved similar graphs. By experimenting with three
classic GNN models on 12 different datasets, we have demonstrated
GraphRetrieval is able to bring substantial improvements to existing GNN models
without comprising the model size and the prediction efficiency. Our work also
firstly validates the feasibility and effectiveness of retrieved-enhanced graph
neural networks.

本文提出一种图检索机制叫做 GraphRetrieval 通过检索训练图来加强现有的图神经网络模型，利用自注意力的适配器从被检索图中获取有益信息并消除噪声，实现了对现有 GNN 模型效果的显著提升。