Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MetaGMT: Improving Actionable Interpretability of Graph Multilinear Networks via Meta-Learning Filtration

Rishabh Bhattacharya, Hari Shankar, Vaishnavi Shivkumar, Ponnurangam Kumaraguru

TL;DR

MetaGMT tackles the risk of spurious explanations in Graph Multilinear Networks by introducing a meta-learning filtration that guides subgraph explanations to be sparse and predictive. The approach uses a bi-level optimization with an inner loop that adapts the classifier on a $k$-hop subgraph using an explanation loss and an outer loop that updates on the full graph based on predictive performance, enabling faithful explanations to generalize. Empirically, MetaGMT improves explanation fidelity (AUC-ROC and Precision@K) and robustness to bias across BA-2Motifs, MUTAG, and SP-Motif, while maintaining competitive accuracy and reducing variance across seeds. This enhances actionable interpretability for GNNs in sensitive domains, supporting debugging, targeted retraining, and human oversight for safer deployment.

Abstract

The growing adoption of Graph Neural Networks (GNNs) in high-stakes domains like healthcare and finance demands reliable explanations of their decision-making processes. While inherently interpretable GNN architectures like Graph Multi-linear Networks (GMT) have emerged, they remain vulnerable to generating explanations based on spurious correlations, potentially undermining trust in critical applications. We present MetaGMT, a meta-learning framework that enhances explanation fidelity through a novel bi-level optimization approach. We demonstrate that MetaGMT significantly improves both explanation quality (AUC-ROC, Precision@K) and robustness to spurious patterns, across BA-2Motifs, MUTAG, and SP-Motif benchmarks. Our approach maintains competitive classification accuracy while producing more faithful explanations (with an increase up to 8% of Explanation ROC on SP-Motif 0.5) compared to baseline methods. These advancements in interpretability could enable safer deployment of GNNs in sensitive domains by (1) facilitating model debugging through more reliable explanations, (2) supporting targeted retraining when biases are identified, and (3) enabling meaningful human oversight. By addressing the critical challenge of explanation reliability, our work contributes to building more trustworthy and actionable GNN systems for real-world applications.

MetaGMT: Improving Actionable Interpretability of Graph Multilinear Networks via Meta-Learning Filtration

TL;DR

MetaGMT tackles the risk of spurious explanations in Graph Multilinear Networks by introducing a meta-learning filtration that guides subgraph explanations to be sparse and predictive. The approach uses a bi-level optimization with an inner loop that adapts the classifier on a -hop subgraph using an explanation loss and an outer loop that updates on the full graph based on predictive performance, enabling faithful explanations to generalize. Empirically, MetaGMT improves explanation fidelity (AUC-ROC and Precision@K) and robustness to bias across BA-2Motifs, MUTAG, and SP-Motif, while maintaining competitive accuracy and reducing variance across seeds. This enhances actionable interpretability for GNNs in sensitive domains, supporting debugging, targeted retraining, and human oversight for safer deployment.

Abstract

The growing adoption of Graph Neural Networks (GNNs) in high-stakes domains like healthcare and finance demands reliable explanations of their decision-making processes. While inherently interpretable GNN architectures like Graph Multi-linear Networks (GMT) have emerged, they remain vulnerable to generating explanations based on spurious correlations, potentially undermining trust in critical applications. We present MetaGMT, a meta-learning framework that enhances explanation fidelity through a novel bi-level optimization approach. We demonstrate that MetaGMT significantly improves both explanation quality (AUC-ROC, Precision@K) and robustness to spurious patterns, across BA-2Motifs, MUTAG, and SP-Motif benchmarks. Our approach maintains competitive classification accuracy while producing more faithful explanations (with an increase up to 8% of Explanation ROC on SP-Motif 0.5) compared to baseline methods. These advancements in interpretability could enable safer deployment of GNNs in sensitive domains by (1) facilitating model debugging through more reliable explanations, (2) supporting targeted retraining when biases are identified, and (3) enabling meaningful human oversight. By addressing the critical challenge of explanation reliability, our work contributes to building more trustworthy and actionable GNN systems for real-world applications.

Paper Structure

This paper contains 19 sections, 9 equations, 1 figure, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. GNN Interpretability
  4. Meta-Learning for Explainability and Graphs
  5. Preliminaries: Graph Multilinear Networks
  6. Subgraph Multilinear Extension (SubMT)
  7. GMT Framework
  8. GMT Variants
  9. MetaGMT: Meta-Filtered GMT
  10. Rationale: Filtration via Meta-Adaptation
  11. Bi-Level Optimization
  12. Experiments
  13. Experimental Setup
  14. Results
  15. Discussion
  16. ...and 4 more sections

Figures (1)

  • Figure 1: The complete pipeline of MetaGMT. We select a node in the graph and extract its k-hop neighbourhood. Then, we use the explainer to find the attention scores. Using the explanation loss defined in \ref{['eq:info_loss']} and \ref{['eq:expl_loss']}, We iteratively train our model (Inner Loop) to predict the correct class. This prediction is compared to ground truth (Outer Loop), and the loss calculated is used to update our explainer and our model.