Fidelity Metric

The most direct fidelity measure: the accuracy of the original GNN when evaluated solely on the explanatory subgraph. A high agreement rate indicates the explainer has captured the structural features that genuinely drive the model's output.

• Computed as: Accuracy(f_original(G_explanation))
• A score of 1.0 means the subgraph perfectly reproduces all original predictions
• Low agreement signals the explainer is highlighting spurious or irrelevant structures

A paired evaluation framework that measures explanation quality through perturbation analysis. This dual metric distinguishes between important and irrelevant substructures.

• Fidelity+: Drop in prediction probability when the most important nodes/edges are removed. Higher is better.
• Fidelity-: Drop when the least important components are removed. Lower is better.
• The difference between these scores reveals how well the explainer discriminates signal from noise

A critical regularization factor that penalizes overly large explanations. True fidelity requires the explanation to be minimal yet sufficient — capturing the causal subgraph without extraneous nodes.

• Measured as the fraction of original graph retained: |E_explanation| / |E_original|
• Optimal sparsity balances fidelity+ against explanation size
• The Graph Information Bottleneck principle formalizes this trade-off mathematically

Measures whether an explanation is unique to the predicted class rather than generically important across all classes. A high-contrastivity explanation changes significantly when the target class changes.

• Evaluated by comparing explanations for different predicted classes on the same input
• Low contrastivity suggests the explainer is identifying dataset-wide artifacts rather than class-specific reasoning
• Critical for multi-class graph classification tasks where structural overlap is common

A robust fidelity metric must remain consistent under small, meaningless input variations. This characteristic tests whether the explanation changes when noise is injected into irrelevant parts of the graph.

• Add random edges or feature noise to non-explanatory regions
• A stable explainer produces nearly identical subgraphs before and after perturbation
• Instability indicates the explainer is overfitting to brittle, non-robust features of the GNN

A fundamental distinction in fidelity evaluation. Faithfulness measures alignment with the model's true internal reasoning, while plausibility measures alignment with human intuition.

• Faithfulness: Does the explanation reflect what the model actually computed?
• Plausibility: Would a human domain expert agree with the explanation?
• High fidelity metrics prioritize faithfulness; a model may use non-intuitive features that are nonetheless correct

A quantitative evaluation score that measures how accurately an explanation subgraph reflects the true reasoning process of the GNN. It is typically assessed by the drop in performance when the explanation is removed.

• Core Principle: If the explanation is faithful, removing it should break the prediction.
• Calculation: Often computed as Accuracy(Original) - Accuracy(Masked).
• Contrast with Fidelity: While fidelity measures how well the explanation mimics the model, faithfulness measures if the explanation captures the model's actual internal logic.

A fidelity assessment method that measures the change in a GNN's prediction after masking or altering the most important nodes or edges identified by an explainer.

• Positive Perturbation: Removing the top-k most important features should cause a sharp decline in prediction score.
• Negative Perturbation: Removing the least important features should have minimal impact.
• Area Under the Curve (AUC): The fidelity curve is often summarized by the AUC of prediction score vs. fraction of features removed.

A principle for learning explainable GNNs by compressing the input graph into a minimal subgraph that retains maximal mutual information about the label, discarding irrelevant structural noise.

• Objective: max I(Y; G_sub) - β * I(G; G_sub) where Y is the label and G_sub is the explanation.
• Key Insight: The bottleneck forces the model to identify the minimal sufficient statistic for prediction.
• Relationship to Fidelity: A subgraph extracted via the information bottleneck should naturally achieve high fidelity.

A self-explainable GNN framework where a generator module extracts a concise, causal subgraph (the rationale) and a predictor module makes a decision based solely on that rationale.

• Cooperative Training: The generator and predictor are trained jointly to maximize prediction accuracy while minimizing rationale size.
• Causal Interpretation: The rationale is treated as the direct cause of the prediction.
• Fidelity Connection: The predictor's output on the rationale serves as the explanation, and fidelity is measured against the original model's behavior.

A method that injects stochasticity into the attention mechanism to automatically select a minimal, label-relevant subgraph for explanation and prediction.

• Mechanism: Learns a Bernoulli distribution over edges, sampling subgraphs during training.
• Regularization: A penalty term controls the size of the extracted subgraph, enforcing sparsity.
• Advantage: Reduces the risk of selecting spurious correlations by introducing randomness, leading to more robust fidelity scores.

A post-hoc explanation method that learns a sparse mask over the messages passed between nodes in each GNN layer to identify which connections are irrelevant to a prediction.

• Layer-Wise Masking: Unlike input-level explanations, GraphMask can identify irrelevant edges at any layer of the GNN.
• Fidelity Evaluation: The prediction is recalculated with masked messages; high fidelity means the masked model's output closely matches the original.
• Key Distinction: Explains the message-passing process, not just the input graph structure.