Self-Explainable Graph Neural Networks

• Self-explainable GNNs are architectures that integrate explanation extraction and prediction to provide innate, human-interpretable rationales.
• They leverage subgraph extraction, prototype reasoning, and concept bottlenecks to manage tasks like node classification, graph classification, and link prediction.
• SE-GNNs enhance transparency, robustness, and trust in graph models while addressing trade-offs in faithfulness, efficiency, and interpretability.

Self-explainable Graph Neural Networks (SE-GNNs) are a family of architectures that tightly couple the prediction and explanation processes within Graph Neural Networks, providing ante-hoc, faithful, and often human-interpretable justifications for their decisions. Unlike post-hoc explainers, SE-GNNs are trained to generate explanations as an integral part of their inference, with the aim of aligning predictive reasoning and interpretability. SE-GNNs encompass a spectrum of mechanisms including subgraph extraction, prototype reasoning, concept bottlenecks, and differentiable logic modules, spanning tasks in node classification, graph classification, link prediction, and dynamic or signed graphs. The study and design of SE-GNNs draws from formal foundations in faithfulness, prime implicant and minimal explanation theory, and aims to bridge the gap between accuracy, robustness and transparency in graph representation learning.

1. Architectural Foundations and Model Taxonomy

The defining characteristic of SE-GNNs is their partition of inference into two (or more) explicit modules: an explanation extractor (often denoted as "detector", "mask generator", or similar) and a predictor (classifier) that conditions its decision strictly on the extracted explanation. The canonical form is

$\hat y = g\left(\,f(G)\,\right)$

where $f$ extracts a subgraph, mask, prototype set, or concept vector from G (graph, subgraph, or node), and $g$ outputs the class label or regression target using only the extracted information (Azzolin et al., 2024, Dai et al., 2021, Huang et al., 2024, Zhang et al., 2021).

Principal SE-GNN classes include:

• Subgraph extractor models: Learn masks or subgraphs (e.g., GSAT, GISST, SMGNN, SES, GCN-SE, GraphOracle), trained via information bottleneck or sparsity priors (Azzolin et al., 2024, Liu et al., 15 Aug 2025, Huang et al., 2024, Fan et al., 2021).
• Prototype-based models: Learn explicit class prototypes as graphs or embedding vectors, with inference by similarity to prototypes (e.g., ProtGNN, ProtEx-GNN, ProtGNN+, PGIB) (Dai et al., 2022, Zhang et al., 2021, Liu et al., 15 Aug 2025).
• Interpretable similarity models: For node and link prediction, explanations are built from interpretable similarity scores and neighbor sets (e.g., SE-GNN for node classification, ILP-GNN for link prediction, SE-SGformer for sign prediction) (Dai et al., 2021, Zhu et al., 2023, Li et al., 2024).
• Concept-bottleneck and logic models: Extract discrete or fuzzy concept assignments, with interpretable logic-based predictors (e.g., CEM/CGN) (Magister et al., 2022).
• Meta-learned/few-shot explainers: Leverage meta-learning to support explainability in data-scarce regimes (e.g., MSE-GNN) (Peng et al., 2024).
• Dual-channel hybrids: Combine SE-GNNs with whitebox rules or other interpretable models for improved OOD generalization (e.g., Dual-Channel GNNs) (Azzolin et al., 4 Feb 2025).

2. Core Algorithms and Explanation Mechanisms

Subgraph and Mask Extraction

Most SE-GNNs for node and graph classification employ a learned subgraph or mask extractor. The detector assigns per-element (node/edge) scores $p_n\in[0,1]$ that, after thresholding or top-K selection, define the explanation subgraph $R$ (Azzolin et al., 2024, Huang et al., 2024). Training objectives typically combine a supervised prediction loss with regularizers (sparsity, entropy minimization, size constraints), or mutual information/bottleneck terms. The predictor operates strictly on $R$, often enforced by removing access to the complement $C = G \setminus R$.

Prototype-based Reasoning

Prototype-based SE-GNNs (e.g., ProtGNN, ProtEx-GNN) learn a library of class prototypes—either as graph embeddings or reconstructed attributed graphs:

• During inference, the test instance is encoded as a graph embedding, then matched by similarity (e.g., negative L2 or cosine) to class prototypes.
• The highest scoring prototype (or a set of top prototypes) constitutes both the model's predictive evidence and the explicit explanation (Dai et al., 2022, Zhang et al., 2021).
• Instance-level explanations arise by matching or overlaying the test graph to the nearest prototypes; class-level explanations are provided directly by the prototype set.

These models routinely incorporate special proto-embedding regularization, cluster/separation losses, initialization from K-means centroids, and prototype realization via graph decoders.

Interpretable Similarity and Neighbor-based Explanation

For node and link-level tasks, SE-GNNs may use an interpretable similarity module:

• Node (or link) predictions are justified by the model's identification of top-K most similar labeled nodes or neighbors, using a combination of local structure and attribute similarity.
• The K-nearest neighbors (for node outputs) or paired neighbor sets (for links) comprise the full explanation, with similarity scores and edge matches directly available (Dai et al., 2021, Zhu et al., 2023, Li et al., 2024).

Concept Bottleneck and Logic-based Explanations

CGN models enforce a “concept bottleneck”:

• Node embeddings are soft-assigned to m-dimensional concept vectors via differentiable softmax assignment and renormalization.
• Predictions are made only from pooled concept vectors, using a shallow logic-explained network (LEN) with sparse attention, yielding DNF-style rules as explanations (Magister et al., 2022).

3. Faithfulness: Theoretical Analysis and Evaluation

Formal Foundations

Faithfulness is central to SE-GNNs and has been formalized in terms of sufficiency and necessity:

• Sufficiency: The extracted explanation R is sufficient if perturbing the complement C does not alter the prediction.
• Necessity: R is necessary if perturbing R itself changes the prediction.
• Unified metrics quantify both, normally via divergences in class probabilities over interventions (masking, zeroing, deletion), and are combined into an overall faithfulness score (Azzolin et al., 2024).

It is also proven that for monolithic (injective) GNNs, strictly faithful explanations are trivial—comprising the full computational graph—so truly informative and non-trivial explanations are possible only in modular (detector–predictor) SE-GNNs that strictly constrain the predictor's information (Azzolin et al., 2024, Azzolin et al., 4 Feb 2025). Architectural features such as Explanation-only Readout (ER), Hard Binary Masks (HS), Content Feature preservation (CF), and Local Aggregation (LA) are found necessary for non-trivial strict faithfulness (Azzolin et al., 2024).

Empirical Metrics and Limitations

Standard faithfulness metrics include:

• Sufficiency/necessity via prediction change upon deletion/masking.
• Unfaithfulness quantified by KL divergence between prediction distributions on the original graph and on the explanation subgraph (Christiansen et al., 2023).

Recent works reveal that different faithfulness metrics can disagree sharply, and some (especially necessity metrics based on random deletion) are insensitive to irrelevant elements in R (Azzolin et al., 2024).

A critical failure mode arises: SE-GNNs can attain optimal predictive accuracy while learning degenerate explanations that encode the label in uninformative or even malicious subgraphs (e.g., anchor nodes, irrelevant features). This can be adversarially induced or occur naturally under strong sparsity or entropy penalties (Azzolin et al., 28 Jan 2026). Most standard faithfulness metrics fail to diagnose such degenerate explanations; robust worst-case sufficiency metrics such as SUFFCAUSE are needed for reliable auditing (Azzolin et al., 28 Jan 2026).

Faithfulness Table

Faithfulness Type	Definition	Limitation
Sufficiency	R contains enough for original label	Trivial in injective/local GNNs
Necessity	Prediction unstable if R perturbed	Insensitive to irrelevant edges
Combined/Harmonic	Harmonic mean of normalized suff/nec	Depends on perturbation choice
Unfaithfulness KL	Divergence between p(G), p(R)	Fails if mapping from concept to graph not bijective
SUFFCAUSE (robust)	Max shift on all supergraphs of R	Computationally costly

4. Class-Level, Instance-Level, and Task-specific Explanations

SE-GNNs can deliver explanations at multiple granularities:

• Class-level: Explicit prototypes, subgraphs, or concepts representing a class. Faithful class-level explanations require that the extracted components hold predictive power for all instances in the class; many current prototype-based models fail this standard due to spurious or unfaithful prototype discovery. Methods such as GraphOracle implement entropy-regularized subgraph selection and dependency modeling to guarantee meaningful class-level explanations, validated by masking-based fidelity tests (Liu et al., 15 Aug 2025).
• Instance-level: The particular prototype matched, subgraph extracted, or neighbor set used for each input instance's prediction.
• Feature-level: Masks or concept activations highlighting relevant node features or attributes (Huang et al., 2024, Magister et al., 2022).
• Temporal/dynamic/heterogeneous graphs: Specialized SE-GNNs (e.g., GCN-SE on dynamic graphs, SE-SGformer for signed graphs) provide explanations in terms of important timesteps or signed neighbor sets (Fan et al., 2021, Li et al., 2024).

5. Training and Optimization Protocols

SE-GNNs are trained with objectives that incorporate both predictive and explanation-oriented losses:

• Explainability losses: Sparsity, entropy, concept-entropy, or mutual information penalties to encourage succinct and interpretable explanations (Huang et al., 2024, Magister et al., 2022).
• Prototype and concept regularization: Cluster/separation/diversity penalties on prototypes or concepts to ensure coverage, uniqueness, and class alignment (Dai et al., 2022, Zhang et al., 2021, Magister et al., 2022).
• Meta-learning: For few-shot SE-GNNs, inner-loop adaptation for fast explainable generalization, with regularizers on rationale size and explanation-fidelity (Peng et al., 2024).
• Dual-channel optimization: Combining a subgraph-extraction SE-GNN with a rule-based channel and an adaptive aggregator, with regularizers that force channel selection where appropriate (Azzolin et al., 4 Feb 2025).
• Instance-level or class-level explanation validation: Masking-based evaluation and entropy penalties to enforce and diagnose faithfulness (Liu et al., 15 Aug 2025, Azzolin et al., 28 Jan 2026).

End-to-end learning and ablation studies strongly support the necessity of tightly coupling explanation extraction with the inference mechanism: decoupled or post-hoc explanation schemes routinely yield explanations that are less faithful, inefficient, or computationally costly (Huang et al., 2024, Zhang et al., 2021, Dai et al., 2022).

6. Practical Applications, Impact, and Limitations

SE-GNNs have achieved demonstrable gains in explainability, auditability, and, in some cases, predictive robustness (including OOD generalization), with state-of-the-art performance on a range of node, graph, link, and dynamic tasks (Dai et al., 2022, Dai et al., 2021, Zhu et al., 2023, Li et al., 2024). Prototypical case studies confirm that SE-GNNs can recover known motifs (e.g., chemical functional groups, subgraph templates), reveal important time steps in dynamic graphs, and output high-purity, compact explanations.

However, fundamental limitations persist:

• Faithfulness at the concept/prototype layer does not guarantee faithfulness at the raw subgraph or node level, due to many-to-many and non-invertible mappings in the grounding step (Christiansen et al., 2023).
• For non-existential, non-local tasks, minimal/prime-implicant explanations can be exponentially large or intractable to enumerate (Azzolin et al., 4 Feb 2025).
• SE-GNNs may be vulnerable to degeneracy, extracting “plausible” yet uninformative explanations—particularly under aggressive sparsity/entropy regularization or in presence of adversarial anchor sets (Azzolin et al., 28 Jan 2026).
• No universally accepted faithfulness metric exists: metrics based on random deletion, sufficiency/necessity, entropy, or KL divergence can disagree and miss key failure cases (Azzolin et al., 2024, Christiansen et al., 2023, Azzolin et al., 28 Jan 2026).
• Training, hyperparameter tuning, and explanation selection criteria (thresholding, top-K, masking budgets) significantly affect explanation quality and interpretability (Huang et al., 2024, Dai et al., 2022, Magister et al., 2022).

7. Open Challenges and Future Directions

Current research on SE-GNNs has established robust foundations for formal explainability, but several open directions remain:

• Faithfulness-completeness tradeoff: Achieving non-trivial, faithful, and succinct explanations especially for non-motif or non-existential tasks (Azzolin et al., 4 Feb 2025).
• Scalable and robust class-level explanations: Efficiently identifying prototypes/subgraphs that generalize globally across all class members, validated by masking-based fidelity (Liu et al., 15 Aug 2025).
• Degeneracy prevention and robust auditing: Designing architectures and faithfulness metrics (e.g., SUFFCAUSE) that preclude degenerate or maliciously-planted explanations and support reliable model auditing (Azzolin et al., 28 Jan 2026).
• Extending to rich graph modalities: Adapting SE-GNNs for dynamic, heterogeneous, signed, and multi-modal graphs; integrating human-in-the-loop corrections and interventions (Magister et al., 2022, Fan et al., 2021, Peng et al., 2024, Li et al., 2024).
• Theoretical characterization and complexity: Understanding separability between minimal explanations, prime implicant explanations, and faithful explanations, along with computational tractability (Azzolin et al., 4 Feb 2025).
• Integration and hybridization: Designing dual- or multi-channel SE-GNNs that blend subgraph, concept, rule-based, and prototype explanations for increased expressivity and robustness (Azzolin et al., 4 Feb 2025, Magister et al., 2022).

The field continues to refine the criteria for trustworthy and actionable SE-GNN explanations, develop efficient architectures for large-scale and domain-critical graphs, and advance principled evaluation protocols that can separate spurious from genuinely causal rationales in graph learning.