New G2I framework generates actionable intervention hypotheses from GNNs

Researchers have developed G2I, a novel two-stage greedy framework designed to generate actionable intervention hypotheses from graph neural networks (GNNs). Unlike existing methods that focus on node-level explanations and continuous optimization, G2I identifies minimal, actionable changes to node features and neighbor conditions for local counterfactuals. For network-level interventions, it frames the problem as a DNF coverage problem, enabling a greedy algorithm with theoretical guarantees. Experiments on synthetic and real-world suicide risk networks show G2I produces scalable, cost-effective strategies that outperform mask-based counterfactual methods. AI

影响 This framework could improve the interpretability and actionability of GNNs in fields like public health and social science.

排序理由 The cluster contains a research paper detailing a new framework for generating intervention hypotheses from graph neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.LG 阅读 →

• Ajitesh Srivastava
• arXiv
• ATP6AP1
• CF-GNNExplainer
• G2I
• graph neural networks
• Hugging Face

• 论文
• 模型发布

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →