Feature-topology cascade perturbation for graph neural network

Hui Cong, Xibei Yang, Keyu Liu, Qihang Guo

Abstract: Graph Neural Network (GNN) has gained great popularity in tackling various analytics tasks focusing on graph data. Data perturbation, particularly feature perturbation, as a promising solution for augmenting graph data, enables GNN to learn powerful representation. However, most feature perturbation strategies excessively emphasize the global perspective, which neglects the contributions of influential nodes from a local perspective. Additionally, the transformed topology corresponding to feature perturbation is insufficiently involved in building networks. To address these issues, we propose a novel plug-and-play architecture, termed Feature-Topology Cascade Perturbation (FTCP) for GNN, which consists of two perturbation stages: celebrity-guided feature perturbation and cascaded topology perturbation. Specifically, on the feature level, we perturb nodes by recognizing celebrities in view of multi-hop structure naturally existing in original topology. This is because figuring out celebrities would be of great help in the representation power of GNN. On the node relationship level, we further track the topology induced by perturbed features via a polarized view, which then assists the original topology to capture richer structure information. Extensive experiments conducted on both regular graph-structure and multi-view data illustrate that our architecture FTCP consistently yields performance improvement when applied to various GNN models.

External IDs:dblp:journals/eaai/CongYLG25