This research paper presents a new method, named SAP, to improve Graph Neural Networks (GNNs). GNNs are effective at learning semantics of graph data. However, the structure information of a graph is usually exploited during pre-training for learning node representations but neglected in the prompt tuning stage. SAP addresses this issue by exploiting structure information in both pre-training and prompt tuning stages. The paper demonstrates the effectiveness of SAP through experiments on node classification and graph classification tasks. SAP also performs well in more challenging few-shot scenarios on both homophilous and heterophilous graphs.
Publication date: 27 Oct 2023
Project Page: https://doi.org/10.1145/nnnnnnn.nnnnnnn
Paper: https://arxiv.org/pdf/2310.17394