GRAFT: Auditing Graph Neural Networks via Global Feature Attribution

2026-05-05 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, medium

Summary

GRAFT is a novel post-hoc global explanation framework introduced on May 5, 2026, by Rishi Raj Sahoo and Subhankar Mishra, designed to interpret Graph Neural Networks (GNNs) by identifying class-level feature importance profiles. While existing GNN explainers focus on structural motifs, GRAFT uniquely explains model behavior globally at the input node attribute level. The method integrates diversity-guided exemplar selection, Integrated Gradients-based attribution, and aggregation to create a comprehensive view of feature influence for each class. It further translates these insights into natural language rules using a self-refining large language model. Evaluated across various datasets, architectures, and experimental setups, GRAFT effectively captures model-relevant features, aids in bias analysis, and supports feature-efficient transfer learning. A structured human evaluation protocol was also developed to assess the interpretability of the generated rules, confirming GRAFT's practical utility in bridging quantitative attribution with human-understandable explanations for GNNs.

Key takeaway

For research scientists developing or deploying Graph Neural Networks, understanding which input features drive predictions is critical for trust and debugging. GRAFT offers a practical solution by providing global, class-level feature importance profiles and translating them into natural language rules. You should consider integrating GRAFT into your GNN development pipeline to enhance model interpretability, facilitate bias detection, and improve transfer learning efficiency.

Key insights

GRAFT provides global, class-level feature importance explanations for GNNs, translating them into natural language rules.

Principles

• Global feature attribution enhances GNN interpretability.
• Combining diverse exemplars improves explanation robustness.
• Natural language rules aid human understanding of GNNs.

Method

GRAFT combines diversity-guided exemplar selection, Integrated Gradients-based attribution, and aggregation to construct class-level feature importance profiles, then uses an LLM with self-refinement to generate natural language rules.

In practice

• Use GRAFT for GNN bias analysis.
• Apply GRAFT for feature-efficient transfer learning.
• Generate human-readable rules for GNN predictions.

Topics

• GRAFT
• Graph Neural Networks
• Feature Attribution
• Model Interpretability
• Node Classification

Code references

• Sennadir/GCORN

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Understanding Convolutions on Graphs — GNNs process graph-structured data by aggregating and combining node features, enabling powerful predictions on complex systems.
• AIMing for Standardised Explainability Evaluation in GNNs: A Framework and Case Study on Graph Kernel Networks — AIM provides a flexible framework for standardized, comprehensive explainability evaluation across diverse GNN models.
• Model Graph Inductive Learning for Knowledge Graph Completion — MGIL improves knowledge graph link prediction by leveraging global structural patterns through a novel model graph and GNN approach.
• Graph Neural Networks for Misinformation Detection: Performance-Efficiency Trade-offs — Graph Neural Networks consistently outperform non-graph methods in misinformation detection with comparable efficiency.
• Beyond Soft Masks: Hard-Perturbation Mixup Explainer for Robust GNN Explainability — HPME uses hard perturbations and structural mixup to generate robust, in-distribution GNN explanations by eliminating irrelevant information.
• Structure and Relationships: Graph Neural Networks and a Pytorch Implementation — GNNs model interconnected data by transforming, aggregating, and updating node features based on neighbor relationships.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.