Graph Structure Learning with Variational Information Bottleneck

Graph Neural Networks (GNNs) have shown promising results on a broad spectrum of applications. Most empirical studies GNNs directly take the observed graph as input, assuming structure perfectly depicts accurate and complete relations between nodes. However, graphs in real-world are inevitably noisy or incomplete, which could even exacerbate quality representations. In this work, we propose novel Variational Information Bottleneck guided Structure Learning framework, namely VIB-GSL, perspective information theory. VIB-GSL is first attempt to advance (IB) principle for learning, providing more elegant universal framework mining underlying task-relevant relations. learns an informative compressive distill actionable specific downstream tasks. deduces variational approximation irregular data form tractable IB objective function, facilitates training stability. Extensive experimental demonstrate that superior effectiveness robustness proposed VIB-GSL.