.png){kind=icon}
.png){kind=icon}
Return
Noise-adaptive bi-directional node-edge flow self-supervised method based Graph Information Bottleneck
.png){kind=icon}
PRE
AI
DOI：10.1016/j.jocs.2026.102911.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Graph Convolutional Networks (GCNs) have driven major advances in graph representation learning, yet existing approaches still rely on unidirectional nodeedge information flow, overlook higher-order global structures, and fall prey to noisy or spurious links that distort embeddings. Moreover, real-world graphs often exhibit dynamic noise patterns and complex node, edge interdependencies, which current models do not jointly capture, leading to brittle performance under perturbations. To address these challenges, we introduce SPSLIB, a Self-Pseudo-Supervised Label Information Bottleneck framework that enforces bidirectional nodeedge exchange via a Dual-Stream Complementary Information Bottleneck, coupled with a Noise-Resilient Learning module and a Noise-Probability-based Labeling scheme for adaptive pseudo-label generation. On benchmarks including MUTAG, PROTEINS, NCI1, DD, OGBG-MOLHIV, Reddit-Threads, SPSLIB achieves up to a 4.5% improvement in classification accuracy under noisy conditions, demonstrating markedly enhanced robustness and representation power.
Keywords:
Graph Convolutional Networks
Information Bottleneck
Node-Edge Interaction
Noise Resilience
Self-Supervised Learning
Journal
J
IF:
3.7
Papers:
205
Citations:
0