Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FSX: Message Flow Sensitivity Enhanced Structural Explainer for Graph Neural Networks

Bizu Feng, Zhimu Yang, Shaode Yu, Zixin Hu

TL;DR

This work tackles the interpretability of Graph Neural Networks by bridging internal computational dynamics with external graph structure. It introduces FSX, a hybrid explainer that first performs flow-sensitivity analysis to identify critical internal message flows during a single forward pass, then maps these flows to a compact key subgraph. Within this subgraph, FSX uses a flow-aware cooperative game with a weighted Shapley-like valuation to fairly assign node contributions, biasing toward coalitions rich in critical information flows. An efficient Monte Carlo approximation confines the game-theoretic valuation to the small subgraph, yielding high-fidelity explanations with reduced runtime. Experiments across multiple datasets and GNN architectures show that FSX outperforms competitive explainers in fidelity while significantly lowering computation time, offering new insights into the structural reasoning behind GNN predictions.

Abstract

Despite the widespread success of Graph Neural Networks (GNNs), understanding the reasons behind their specific predictions remains challenging. Existing explainability methods face a trade-off that gradient-based approaches are computationally efficient but often ignore structural interactions, while game-theoretic techniques capture interactions at the cost of high computational overhead and potential deviation from the model's true reasoning path. To address this gap, we propose FSX (Message Flow Sensitivity Enhanced Structural Explainer), a novel hybrid framework that synergistically combines the internal message flows of the model with a cooperative game approach applied to the external graph data. FSX first identifies critical message flows via a novel flow-sensitivity analysis: during a single forward pass, it simulates localized node perturbations and measures the resulting changes in message flow intensities. These sensitivity-ranked flows are then projected onto the input graph to define compact, semantically meaningful subgraphs. Within each subgraph, a flow-aware cooperative game is conducted, where node contributions are evaluated fairly through a Shapley-like value that incorporates both node-feature importance and their roles in sustaining or destabilizing the identified critical flows. Extensive evaluation across multiple datasets and GNN architectures demonstrates that FSX achieves superior explanation fidelity with significantly reduced runtime, while providing unprecedented insights into the structural logic underlying model predictions--specifically, how important sub-structures exert influence by governing the stability of key internal computational pathways.

FSX: Message Flow Sensitivity Enhanced Structural Explainer for Graph Neural Networks

TL;DR

This work tackles the interpretability of Graph Neural Networks by bridging internal computational dynamics with external graph structure. It introduces FSX, a hybrid explainer that first performs flow-sensitivity analysis to identify critical internal message flows during a single forward pass, then maps these flows to a compact key subgraph. Within this subgraph, FSX uses a flow-aware cooperative game with a weighted Shapley-like valuation to fairly assign node contributions, biasing toward coalitions rich in critical information flows. An efficient Monte Carlo approximation confines the game-theoretic valuation to the small subgraph, yielding high-fidelity explanations with reduced runtime. Experiments across multiple datasets and GNN architectures show that FSX outperforms competitive explainers in fidelity while significantly lowering computation time, offering new insights into the structural reasoning behind GNN predictions.

Abstract

Despite the widespread success of Graph Neural Networks (GNNs), understanding the reasons behind their specific predictions remains challenging. Existing explainability methods face a trade-off that gradient-based approaches are computationally efficient but often ignore structural interactions, while game-theoretic techniques capture interactions at the cost of high computational overhead and potential deviation from the model's true reasoning path. To address this gap, we propose FSX (Message Flow Sensitivity Enhanced Structural Explainer), a novel hybrid framework that synergistically combines the internal message flows of the model with a cooperative game approach applied to the external graph data. FSX first identifies critical message flows via a novel flow-sensitivity analysis: during a single forward pass, it simulates localized node perturbations and measures the resulting changes in message flow intensities. These sensitivity-ranked flows are then projected onto the input graph to define compact, semantically meaningful subgraphs. Within each subgraph, a flow-aware cooperative game is conducted, where node contributions are evaluated fairly through a Shapley-like value that incorporates both node-feature importance and their roles in sustaining or destabilizing the identified critical flows. Extensive evaluation across multiple datasets and GNN architectures demonstrates that FSX achieves superior explanation fidelity with significantly reduced runtime, while providing unprecedented insights into the structural logic underlying model predictions--specifically, how important sub-structures exert influence by governing the stability of key internal computational pathways.
Paper Structure (16 sections, 12 equations, 4 figures, 2 algorithms)

This paper contains 16 sections, 12 equations, 4 figures, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related work
  3. Graph Neural Networks
  4. Explainability in Graph Neural Networks
  5. Method
  6. Flow-Sensitivity Analysis via Localized Perturbation
  7. Flow-Aware Subgraph Identification and Valuation
  8. From Flows to Subgraphs: Distilling Causal Pathways
  9. Flow-Aware Cooperative Game Formulation: Valuing Structural Stability
  10. Efficient Weighted Shapley-like Value Approximation: Bias for Causal Fidelity
  11. Experiment
  12. Datasets and Baselines
  13. Settings and evaluation metrics
  14. Explanation Comparison
  15. Efficiency Analysis
  16. ...and 1 more sections

Figures (4)

  • Figure 1: An illustration of FSX
  • Figure 2: Fidelity+ values were evaluated on 4 datasets using GINs under different sparsity levels, where a higher Fidelity+ value indicates better performance.
  • Figure 3: Fidelity- values were evaluated on 4 datasets using GINs under different sparsity levels, where a lower Fidelity- value indicates better performance.
  • Figure 4: Average Time