Self-Interpretable Graph Neural Networks

• SI-GNNs are graph models that integrate built-in, faithful rationales with predictive objectives during the forward pass.
• They employ mechanisms like attention distributions, subgraph masks, prototypes, and symbolic rules to elucidate predictions in diverse tasks.
• SI-GNNs improve interpretability over post-hoc methods, though challenges remain in scalability, hyperparameter tuning, and expressivity.

Self-interpretable graph neural networks (SI-GNNs) are graph models whose architectures are designed to produce faithful, built-in rationales concurrently with predictions. Unlike post-hoc explainers, which probe a pre-trained black-box after training, SI-GNNs embed interpretable mechanisms—attention distributions, subgraph masks, concept bottlenecks, prototypes, or symbolic rules—directly in the forward pass, so explanations are produced simultaneously with predictions during inference (Ji et al., 26 Jan 2025). Across the literature, the explanatory object varies with the task: node classification can be justified by labeled exemplars and matched local structures, graph classification by sparse subgraphs or prototype graphs, link prediction by pair-specific neighbor sets, and symbolic models by explicit predicates and quantitative disjunctive normal form rules (Dai et al., 2021, Zhu et al., 2023, Geng et al., 18 Feb 2026).

1. Definition and scope

The central distinction between SI-GNNs and post-hoc explainers is architectural rather than cosmetic. In SI-GNNs, the explanatory variables are optimized jointly with the predictive objective and directly affect the model’s computation; in post-hoc explanation, the model first predicts and an auxiliary procedure later attempts to rationalize that prediction (Ji et al., 26 Jan 2025). Early formulations made this contrast explicit. GISST learns task-aligned importance probabilities for both topology and node features and uses them to weight the actual inputs of any downstream GNN, thereby yielding interpretable subgraphs and feature subsets intrinsically tied to task performance (Lin et al., 2020). SE-GNN studies self-explainable node classification by retrieving $K$ nearest labeled nodes under an interpretable similarity that combines node similarity and local-structure similarity, so the retrieved exemplars are not auxiliary commentary but the decision basis itself (Dai et al., 2021).

As the field broadened, the unit of explanation became task-dependent. ILP-GNN argues that link prediction cannot be explained by simply combining two node explanations, because a predicted link depends on pair-specific factors jointly relevant to both endpoints; it therefore returns two $K$-neighbor sets, one for each node in the pair, together with normalized importance weights (Zhu et al., 2023). Graph-level methods likewise moved beyond purely local rationales. GIP argues that graph classification often depends on long-range dependencies and global interactions rather than only small salient subgraphs, and therefore explains predictions through learnable global interactive patterns defined on coarsened graphs (Wang et al., 2024). More recently, GraphOracle distinguishes instance-level from class-level self-explainability and treats the discovery of class-discriminative subgraphs as a separate problem, with masking-based evaluation of whether those subgraphs truly govern predictions for an entire class (Liu et al., 15 Aug 2025).

2. Methodological families

A survey of self-interpretable neural networks organizes graph-specific SI-GNNs into five perspectives: attribution-based, function-based, concept-based, prototype-based, and rule-based (Ji et al., 26 Jan 2025).

Perspective	Explanation object	Representative examples
Attribution-based	Edge/node probabilities, stochastic gates, masks	GISST (Lin et al., 2020), SES (Huang et al., 2024), i-WiViG (Obadic et al., 11 Mar 2025), SIG (Fang et al., 2024)
Function-based / subgraph-selection	Sparse or connected explanatory subgraphs used in forward computation	L2XGNN (Serra et al., 2022), GMT (Chen et al., 2024), MSInterpreter (Li et al., 2023)
Concept-based	Motifs, temporal patterns, meta-paths, logical concepts	TempME, TECHS, HGCN (Ji et al., 26 Jan 2025)
Prototype-based	Exemplar graphs, subgraph prototypes, global interactive patterns	ProtGNN (Zhang et al., 2021), prototype-based SI-GNN (Dai et al., 2022), GIP (Wang et al., 2024), GraphOracle (Liu et al., 15 Aug 2025)
Rule-based / symbolic	Logical clauses, predicates, Q-DNF rules	SymGraph (Geng et al., 18 Feb 2026), TECHS (Ji et al., 26 Jan 2025)

These families are not mutually exclusive. Prototype methods often incorporate subgraph extraction or clustering; attribution-based models may use stochastic bottlenecks that function as concept selectors; symbolic methods can be viewed as an extreme form of function-based transparency in which the explanatory mechanism is a rule set rather than a mask or prototype. What unifies them is the requirement that the explanatory object be part of the predictive pipeline rather than a later reconstruction (Ji et al., 26 Jan 2025).

3. Architectural mechanisms

One dominant mechanism is intrinsic masking or gating. GISST introduces edge-importance probabilities $P_{A_s}$ and feature-importance probabilities $P_{X_s}$ and feeds the weighted inputs $\Phi(A \odot P_{A_s}, X \odot P_{X_s})$ into an arbitrary downstream GNN, with entropy and $L_1$ regularization encouraging sparse, near-binary explanations (Lin et al., 2020). SES similarly co-trains a global mask generator with a two-layer encoder to produce feature masks $M_f$ and structure masks $M_s$ during training; those masks are then reused to construct positive and negative neighborhoods for a triplet loss that explicitly feeds explanation signals back into representation learning (Huang et al., 2024). In i-WiViG, window-based graph processing preserves local correspondence between nodes and image regions, while a GSAT-based bottleneck learns edge probabilities for long-range window relations and exposes the explanatory subgraph directly at inference time (Obadic et al., 11 Mar 2025). SIG adapts the same intrinsic logic to continuous-time dynamic graphs by extracting temporal and structural causal subgraphs and coupling them with interventional heads that implement do-operations against learned confounder dictionaries (Fang et al., 2024).

A second mechanism is explicit subgraph or neighbor selection. L2XGNN learns a discrete explanatory subgraph for each input graph and then runs a standard message-passing GNN exclusively on that subgraph, with sparsity and optional connectivity enforced as hard feasibility constraints rather than soft penalties (Serra et al., 2022). MSInterpreter performs a related operation at the level of message-passing paths: it computes edge scores from structural overlap and one-layer node embeddings, sorts those scores, and replaces the original adjacency with a selected critical-path adjacency before any GNN layer aggregates messages (Li et al., 2023). For node classification, SE-GNN retrieves the $K$ nearest labeled nodes according to a convex combination of node-feature similarity and local-structure similarity; for link prediction, ILP-GNN performs hard, pair-specific Top-$K$ neighbor selection with common-neighbor priority and then aggregates the selected neighbors into pair-specific embeddings (Dai et al., 2021, Zhu et al., 2023). GMT formalizes this family theoretically through the subgraph multilinear extension (SubMT) and argues that faithful interpretable subgraph learning requires approximating the expectation over discrete subgraphs rather than merely feeding a soft graph into the classifier (Chen et al., 2024).

Prototype and symbolic methods shift the explanatory object from masks to reusable structures. ProtGNN compares each graph embedding to a small set of learned class-specific prototypes and explains predictions by case-based reasoning; ProtGNN+ adds a conditional subgraph sampler that identifies which part of the input graph best matches each prototype (Zhang et al., 2021). A related prototype-based SI-GNN generates explicit prototype graphs from learnable node embeddings, classifies by similarity to those graphs, and uses the matched prototypes as class-level and instance-level explanations (Dai et al., 2022). GIP learns graph-level interactive pattern prototypes after constrained graph coarsening, while GraphOracle learns a bank of random-walk subgraphs and class-specific graph–subgraph–prediction dependencies whose final linear weights determine class-level explanations (Wang et al., 2024, Liu et al., 15 Aug 2025). SymGraph replaces message passing altogether with structural hashing, orbit-aware aggregation, and local predicates $K$0, and then predicts with quantitative disjunctive normal form rules over predicate counts (Geng et al., 18 Feb 2026).

4. Faithfulness, evaluation, and recurring controversies

Faithfulness is the principal evaluation criterion of SI-GNNs, but it is operationalized differently across tasks. ILP-GNN evaluates pairwise explanations by deletion tests and aggregated $K$1, reporting larger AUC drops than GAT and CONPI when removing the top-ranked neighbors (Zhu et al., 2023). i-WiViG evaluates explanation quality through insertion AUC, post-hoc infidelity, and sparsity, and reports that its intrinsic subgraphs reduce the infidelity of post-hoc explanations without sacrificing explanation sparsity (Obadic et al., 11 Mar 2025). SIG evaluates explanations on continuous-time dynamic graphs with Fidelity$K$2, Sparsity, and AUFSC, and shows large gains over post-hoc explainers that frequently exceed a 24-hour time limit on full datasets (Fang et al., 2024). GraphOracle introduces Fidelity$K$3, Fidelity$K$4, and Fidelity$K$5 for class-level explanations and reports that on MUTAG, at 0.50 sparsity, Fidelity$K$6 is 32.74% for GraphOracle versus 0.50% for ProtGNN and $K$7 for PGIB (Liu et al., 15 Aug 2025).

Several controversies recur in this literature. One concerns whether soft attention or soft masks can stand in for the distribution over discrete explanatory subgraphs. GMT answers negatively in general: it argues that many attention-based interpretable GNNs can have a huge gap in fitting SubMT, and relates better SubMT approximation to stronger counterfactual fidelity (Chen et al., 2024). A second concerns explanation granularity. ILP-GNN shows that node explainers do not transfer cleanly to pair-specific link explanations, while GIP argues that graph-level tasks often require global interactions rather than only local subgraphs (Zhu et al., 2023, Wang et al., 2024). A third concerns the distinction between instance-level and class-level self-explainability. GraphOracle’s masking-based evaluation concludes that prototype-based methods such as ProtGNN and PGIB do not deliver effective class-level explanations, even when their instance-level explanations are plausible (Liu et al., 15 Aug 2025). These debates indicate that “intrinsic explanation” is not a single property but a family of constraints whose adequacy depends on the target task and explanatory granularity.

5. Tasks and empirical landscape

On node-level tasks, SI-GNNs typically combine competitive prediction with explicit local rationales. SE-GNN achieves the best reported accuracy among compared methods on Citeseer and Pubmed, with 73.8±0.6 and 80.0±0.2 respectively, while also achieving markedly higher human evaluation scores for its exemplar-based explanations than MLP-K, GCN-K, or GIN-K (Dai et al., 2021). SES reports 90.64±0.65 on Cora with a GCN backbone and explanation AUC of 99.8 on BA-Shapes, and its ablations show that removing the triplet loss produces the largest drop in accuracy, which is consistent with its premise that explanations can be reused to improve prediction (Huang et al., 2024).

On graph classification, subgraph-selection and prototype methods dominate. L2XGNN matches or slightly improves the accuracy of full-graph GNNs while making predictions from sparse explanatory subgraphs only, and it generates explanations far faster at test time than SubgraphX or GNNExplainer (Serra et al., 2022). ProtGNN and ProtGNN+ achieve accuracy on par with or better than their backbones across molecules, text graphs, and synthetic benchmarks; for example, on BBBP with a GCN backbone, the original model reports 84.6%±3.4, ProtGNN 89.4%±4.1, and ProtGNN+ 88.0%±4.6 (Zhang et al., 2021). GIP reports significantly superior interpretability metrics such as consistency and silhouette score while maintaining competitive graph classification performance, particularly on synthetic datasets where global interaction patterns are ground truth (Wang et al., 2024). SymGraph pushes this trajectory further by presenting a symbolic alternative that reports state-of-the-art performance together with 10×–100× CPU-time speedups and fine-grained rules aligned with domain formalisms such as SMARTS (Geng et al., 18 Feb 2026).

Link prediction and temporal reasoning expose different explanatory requirements. ILP-GNN reports top results among tested methods on Ogbn-arxiv and Ogbl-collab, with AUC 95.42±0.03 and 97.17±0.01, while simultaneously returning two weighted $K$8-neighbor sets as pair-specific rationales (Zhu et al., 2023). SIG, designed for continuous-time dynamic graphs, reports AP/AUC of 99.94/99.94 on Wikipedia and 99.99/99.99 on Reddit, together with strong explanation fidelity and robustness to shortcut features under synthetic OOD interventions (Fang et al., 2024).

The SI-GNN paradigm has also spread beyond conventional graph mining. i-WiViG applies intrinsic subgraph reasoning to remote sensing, reporting Accuracy/F1 of 0.91/0.91 on RESISC45 and $K$9 of 0.46/0.50 on Liveability while exposing long-range window relations as explanations (Obadic et al., 11 Mar 2025). IGNNet reformulates each tabular instance as a feature graph and constrains the final prediction to be an additive sum of per-feature contributions; it reports performance on par with XGBoost, Random Forests, and TabNet, and states that its explanations are aligned with the true Shapley values of the features without any additional computational overhead (Alkhatib et al., 2023).

6. Limitations and research directions

Despite rapid diversification, SI-GNNs still face several unresolved limitations. Scalability is a recurring concern: ProtGNN’s prototype projection relies on Monte Carlo Tree Search, ILP-GNN’s pair-specific neighbor scoring scales with neighborhood sizes, i-WiViG notes that very high-resolution imagery may require hierarchical windowing, and L2XGNN incurs training overhead from perturb-and-MAP sampling and implicit MLE (Zhang et al., 2021, Zhu et al., 2023, Obadic et al., 11 Mar 2025, Serra et al., 2022). Hyperparameter sensitivity is equally common. SES depends on the balance between explanation and prediction losses, i-WiViG reports that regression explanations are sensitive to the GSAT sparsity prior $P_{A_s}$0, and ILP-GNN notes sensitivity to $P_{A_s}$1, $P_{A_s}$2, and the structure-feature trade-off $P_{A_s}$3 (Huang et al., 2024, Obadic et al., 11 Mar 2025, Zhu et al., 2023).

Formal guarantees remain uneven. ProtGNN provides a theorem giving conditions under which ProtGNN+ preserves ProtGNN’s decision, and GMT provides a theoretical framework linking SubMT approximation to counterfactual fidelity, but several influential systems explicitly state that they do not provide formal guarantees on faithfulness, robustness, or identifiability (Zhang et al., 2021, Chen et al., 2024, Zhu et al., 2023). This gap is one reason the survey highlights standardization of evaluation protocols, unified graph explanation benchmarks, human-in-the-loop studies, and completeness metrics as open challenges for SI-GNNs (Ji et al., 26 Jan 2025).

A final research frontier concerns the relation between self-interpretability and the expressivity limits of message passing. SymGraph argues that many self-explainable GNNs inherit the 1-Weisfeiler–Lehman barrier from message-passing backbones and therefore offer only coarse-grained explanations of structures that domain experts would describe more compositionally (Geng et al., 18 Feb 2026). GraphOracle, meanwhile, shows that instance-level explanations are not a reliable proxy for class-level faithfulness, and SIG shows that shortcut-robust explanations on dynamic graphs require explicit causal treatment of confounders rather than only sparse masking (Liu et al., 15 Aug 2025, Fang et al., 2024). Taken together, these results place SI-GNNs at the intersection of interpretability, causal invariance, scalable structured reasoning, and expressive graph representation learning.