Research map / Graph learning / GNN architectures

Reconsidering Faithfulness in Regular, Self-Explainable and Domain Invariant GNNs

ICLR 2025 · Graph learning › GNN architectures · approach: GNN explainability

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Analyzes faithfulness metrics for regular and modular GNN explanations, showing metrics are not interchangeable and necessity metrics can be insensitive to irrelevant edges, that strictly faithful explanations are trivial for injective GNNs, and that faithfulness relates to domain-invariant OOD generalization.

Problem

Faithfulness of GNN explanations is measured by many non-interchangeable metrics that can be insensitive to explanation size, and it is unclear whether optimizing faithfulness is desirable for regular, self-explainable, and domain-invariant GNNs.

Approach

Unifies existing sufficiency and necessity faithfulness metrics under one template (divergence d plus interventional distributions P_R and P_C), proposes a necessity estimator with uniform/budgeted interventions that depends on irrelevant edge count, and proves strictly faithful explanations are trivial for injective regular GNNs but attainable for modular GNNs under specific design strategies.

Key results

• Strictly faithful explanations are trivial (fully determined by the input computational graph) for injective regular GNNs, per proposition on node and graph classification.
• Augmenting SE-GNNs with the ER strategy improved faithfulness over the base model in 13 of 21 cases and was on par in 4.
• Faithfulness and likelihood difference between ID and OOD splits show a statistically significant anti-correlation (Pearson -0.83, p=0.01) for DI-GNNs.

Measured effects reported in the paper

Surprising findings

• For injective regular GNNs, perfectly faithful explanations are completely uninformative.
• Robust fidelity minus and fidelity plus are systematically insensitive to the number of truly irrelevant edges in the explanation.

Limitations

• Hard-score and locality strategies often degrade accuracy and yield inconsistent faithfulness gains.
• Necessity metric depends on budget hyperparameter choice, which can bias the estimate.
• Empirical results limited to seven graph datasets and six modular architectures.

Experimental setup

Datasets: BaMS Motif2-Basis Motif-Size Motif-Basis BBBP CMNIST LBAPcore SST2

Benchmarks: BaMS Motif2-Basis Motif-Size BBBP CMNIST LBAPcore SST2

Metrics: accuracy faithfulness (harmonic mean of normalized sufficiency and ne OOD accuracy Pearson correlation

Baselines: fidelity minus robust fidelity minus fidelity plus probability of sufficiency unfaithfulness

Builds on: GSAT LECI CIGA RAGE GISST

Closest papers by meaning

• Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks (ICLR 2024 · similarity 0.93)
• Beyond Topological Self-Explainable GNNs: A Formal Explainability Perspective (ICML 2025 · similarity 0.92)
• Self-Interpretable Graph Learning with Sufficient and Necessary Explanations (AAAI 2024 · similarity 0.91)
• Redundancy Undermines the Trustworthiness of Self-Interpretable GNNs (ICML 2025 · similarity 0.91)
• Beyond Weisfeiler-Lehman: A Quantitative Framework for GNN Expressiveness (ICLR 2024 · similarity 0.90)
• GNNX-BENCH: Unravelling the Utility of Perturbation-based GNN Explainers through In-depth Benchmarking (ICLR 2024 · similarity 0.89)