Discrete Curvature Graph Information Bottleneck

• DOI: 10.1609/aaai.v39i16.33831
• Publication Date: 2025-04-11T12:42:50Z
• Authors: Xingcheng Fu; Jian Wang; Yisen Gao; Qingyun Sun; Haonan Yuan; Jianxin Li; Xianxian Li
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Machine Learning (cs.LG)

Graph neural networks(GNNs) have been demonstrated to depend on whether the node effective information is sufficiently passing. Discrete curvature (Ricci curvature) used study graph connectivity and propagation efficiency with a geometric perspective, has raised in recent years explore efficient message-passing structure of GNNs. However, most empirical studies are based directly observed structures or heuristic topological assumptions, lack in-depth exploration underlying optimal transport for downstream tasks. We suggest that optimization more essential than rewiring learning richer characterization better interpretability. From both geometry theory perspectives, we propose novel Curvature Information Bottleneck (CurvGIB) framework optimize learn representations simultaneously. CurvGIB advances Variational (VIB) principle Ricci pattern specific The learned refine graph, representation fully efficiently learned. Moreover, computational complexity differentiation, combine flow VIB deduce approximation form tractable IB objective function. Extensive experiments various datasets demonstrate superior effectiveness interpretability CurvGIB.