Graph-based Integrated Gradients for Explaining Graph Neural Networks

2025-09-09

AI Summary

Integrated Gradients (IG), a prominent attribution method, relies on continuity assumptions and thus cannot be directly applied to discrete graph structures. Method: This paper introduces GB-IG—the first IG-adaptation framework for Graph Neural Networks (GNNs)—which extends the IG paradigm to the graph domain by defining topology-aware gradient integration paths and jointly approximating integrals over both node features and structural elements (e.g., edge weights and adjacency relations). Contribution/Results: Extensive experiments on four synthetic and three real-world graph datasets demonstrate that GB-IG significantly outperforms existing baselines in identifying salient nodes, edges, and substructures. It improves explanation fidelity and human interpretability in node classification tasks. By rigorously bridging the theoretical and practical gap of IG on discrete graph data, GB-IG establishes a novel paradigm for GNN interpretability.

Technology Category

Machine Learning: Graph-based Machine LearningReasoning under Uncertainty: Graphical ModelsComputer Vision: Interpretability, Explainability, and Transparency

Application Category

Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphsSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsSearch and Retrieval-Augmented AI: Web query analysis, representation and understanding

Abstract

Integrated Gradients (IG) is a common explainability technique to address the black-box problem of neural networks. Integrated gradients assumes continuous data. Graphs are discrete structures making IG ill-suited to graphs. In this work, we introduce graph-based integrated gradients (GB-IG); an extension of IG to graphs. We demonstrate on four synthetic datasets that GB-IG accurately identifies crucial structural components of the graph used in classification tasks. We further demonstrate on three prevalent real-world graph datasets that GB-IG outperforms IG in highlighting important features for node classification tasks.

Problem

• Extending Integrated Gradients to discrete graph structures
• Identifying crucial structural components in graph classification
• Improving feature importance highlighting for node classification tasks

Innovation

• Extends Integrated Gradients to graph structures
• Identifies crucial structural components in graphs
• Outperforms original IG on node classification tasks