[3E1-GS-2d] Machine learning

9:00 AM - 9:15 AM JST(12:00 AM - 12:15 AM UTC)

[3E1-GS-2d-01] Structural Regularization Learning of Graph Neural Networks based on Conditional Independency

〇Sawaka Wada¹, Sugahara Shouta¹, Maomi Ueno¹ (1. The University of Electro-Communications)

Keywords:

Graph Neural Networks,Graph Structure Learning

Recently, Graph Structure Learning (GSL) has emerged as an effective framework for Graph Neural Networks (GNN). GSL learns an adjacency matrix from data. This is useful when the underlying graph structure is unknown or incomplete. However, conventional task-driven GSL methods often absorb observational noise into the learned structure. This happens because they minimize prediction loss directly. As a result, generalization degrades and interpretability becomes poor.To address this issue, we propose the CI-gated GNN. It introduces a Conditional Independence (CI) structure matrix as an edge-wise gate. This gating suppresses message passing that is inconsistent with CI while emphasizing connections that indicate strong dependencies. In classification experiments on seven UCI datasets, our method outperformed existing GSL frameworks (IDGL, GAug, and LDS) and an ablation model in terms of AUC. Furthermore, in structure recovery tasks with synthetic SEM data, it achieved higher F1 scores than DAGMA and NOTEARS. These results demonstrate that incorporating CI-based statistical principles into task-driven structure learning improves both generalization and interpretability.