Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On the Feasibility of Fidelity$^-$ for Graph Pruning

Yong-Min Shin, Won-Yong Shin

TL;DR

The paper investigates whether fidelity^- can serve as a practical criterion for graph pruning by proposing Fidelity^- inspired Pruning (FiP), which aggregates local edge attributions into a global edge mask $G^{\phi}$ to remove edges and improve GNN efficiency. It demonstrates, across four datasets and seven attribution methods, that general XAI techniques (e.g., Att, SA, IG) often outperform GNN-tailored explanations for pruning, challenging assumptions about specialized explanations being best for fidelity-based tasks. The study also shows that simple aggregation (sum vs average) yields similar pruning performance, and that fidelity^- scores do not always align with pruning effectiveness, highlighting both the promise and limitations of fidelity-guided pruning. Overall, FiP provides a lightweight, interpretable approach to sparsifying graphs, with potential to reduce computational costs while producing sparser, human-friendly explanations; future work includes more sophisticated aggregation schemes and deeper exploration of sparsity vs interpretability.

Abstract

As one of popular quantitative metrics to assess the quality of explanation of graph neural networks (GNNs), fidelity measures the output difference after removing unimportant parts of the input graph. Fidelity has been widely used due to its straightforward interpretation that the underlying model should produce similar predictions when features deemed unimportant from the explanation are removed. This raises a natural question: "Does fidelity induce a global (soft) mask for graph pruning?" To solve this, we aim to explore the potential of the fidelity measure to be used for graph pruning, eventually enhancing the GNN models for better efficiency. To this end, we propose Fidelity$^-$-inspired Pruning (FiP), an effective framework to construct global edge masks from local explanations. Our empirical observations using 7 edge attribution methods demonstrate that, surprisingly, general eXplainable AI methods outperform methods tailored to GNNs in terms of graph pruning performance.

On the Feasibility of Fidelity$^-$ for Graph Pruning

TL;DR

The paper investigates whether fidelity^- can serve as a practical criterion for graph pruning by proposing Fidelity^- inspired Pruning (FiP), which aggregates local edge attributions into a global edge mask to remove edges and improve GNN efficiency. It demonstrates, across four datasets and seven attribution methods, that general XAI techniques (e.g., Att, SA, IG) often outperform GNN-tailored explanations for pruning, challenging assumptions about specialized explanations being best for fidelity-based tasks. The study also shows that simple aggregation (sum vs average) yields similar pruning performance, and that fidelity^- scores do not always align with pruning effectiveness, highlighting both the promise and limitations of fidelity-guided pruning. Overall, FiP provides a lightweight, interpretable approach to sparsifying graphs, with potential to reduce computational costs while producing sparser, human-friendly explanations; future work includes more sophisticated aggregation schemes and deeper exploration of sparsity vs interpretability.

Abstract

As one of popular quantitative metrics to assess the quality of explanation of graph neural networks (GNNs), fidelity measures the output difference after removing unimportant parts of the input graph. Fidelity has been widely used due to its straightforward interpretation that the underlying model should produce similar predictions when features deemed unimportant from the explanation are removed. This raises a natural question: "Does fidelity induce a global (soft) mask for graph pruning?" To solve this, we aim to explore the potential of the fidelity measure to be used for graph pruning, eventually enhancing the GNN models for better efficiency. To this end, we propose Fidelity-inspired Pruning (FiP), an effective framework to construct global edge masks from local explanations. Our empirical observations using 7 edge attribution methods demonstrate that, surprisingly, general eXplainable AI methods outperform methods tailored to GNNs in terms of graph pruning performance.
Paper Structure (21 sections, 3 figures, 2 tables)

This paper contains 21 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Fidelity.
  4. Graph pruning.
  5. Methodology
  6. Basic Notations and Problem Settings
  7. Fidelity$^{-}$-inspired Pruning Framework
  8. Empirical Observations
  9. Basic Settings
  10. Explanation Methods
  11. Attention (Att).
  12. Saliency (SA).
  13. Integrated Gradient (IG).
  14. Guided Backpropagation (GB).
  15. GNNExplainer (GNNEx).
  16. ...and 6 more sections

Figures (3)

  • Figure 1: The overview of the FiP framework.
  • Figure 2: Graph pruning performance of FiP for 7 edge attribution methods as well as a random baseline on 4 benchmark datasets. The grey area indicates the performance of random attributions, and the dashed line indicate the test performance without any pruning.
  • Figure 3: Visualizations of graph pruning using different edge attribution methods by removing 50% of the edges from the original graph.