DiGNNExplainer: A Unified Framework for GNN Explanation

• DiGNNExplainer is a research umbrella for explaining GNNs in dynamic, directed, and multi-relational settings with flexible, compact subgraph recovery.
• It leverages variants like DGExplainer, DRExplainer, and TAGE to address temporal dependencies, edge directionality, and task-agnostic challenges.
• Empirical results demonstrate improved fidelity, sparsity, and runtime efficiency across diverse datasets, providing practical insights for dynamic graph analysis.

Searching arXiv for papers relevant to “DiGNNExplainer,” including dynamic and directed GNN explainers. DiGNNExplainer denotes, in current arXiv usage, not a single canonical algorithm but a cluster of explainer designs for graph neural networks in settings where plain static, undirected, single-task assumptions are inadequate. In the available literature, the term is explicitly aligned with DGExplainer for dynamic GNNs, is used as a natural abstraction of DRExplainer’s directed mask-learning framework, and is also invoked to describe TAGE’s task-agnostic, embedding-centric regime. This suggests that DiGNNExplainer is best understood as a research umbrella for explanation mechanisms that preserve temporal dependence, edge directionality, multi-relational structure, or downstream-task flexibility while still seeking compact, faithful explanatory subgraphs or masks (Xie et al., 2022, Shi et al., 2024, Xie et al., 2022).

1. Terminological scope

The label is used or implied in several adjacent senses rather than as the title of one standardized method.

Sense	Representative paper	Core formulation
Dynamic GNN explanation	DGExplainer (Xie et al., 2022)	Relevance back-propagation through a GCN–GRU over graph snapshots
Directed GNN explanation	DRExplainer (Shi et al., 2024)	Mask learning on directed multi-relational adjacency tensors
Task-agnostic GNN explanation	TAGE (Xie et al., 2022)	Self-supervised embedding explainer conditioned by downstream dimension importance

In the dynamic setting, the term is mapped directly to DGExplainer, which explains predictions on a sequence of graph snapshots by propagating relevance backward through temporal and spatial modules. In the directed setting, DRExplainer is presented as a concrete template from which a general directed-GNN explainer would learn continuous masks on directed adjacency tensors to preserve target outputs. In the task-agnostic setting, TAGE is described as a “DiGNNExplainer” in the sense that it explains the encoder rather than a fixed downstream head and is reused across unseen tasks (Xie et al., 2022, Shi et al., 2024, Xie et al., 2022).

2. Foundational formulation in post-hoc GNN explanation

The conceptual baseline for nearly all DiGNNExplainer-style methods is GNNExplainer, which introduced the general post-hoc program of identifying a compact explanatory subgraph and a small subset of node features for a trained GNN prediction. Its core objective is mutual-information maximization:

$$\max_{G_S, F} MI\left(Y, (G_S, F)\right) = H(Y) - H(Y \mid G = G_S, X = X_S^F),$$

with a continuous relaxation of the explanatory subgraph through a learnable edge mask $M$ and masked adjacency

$\tilde{A} = A_c \odot \sigma(M).$

The practical loss combines prediction preservation with size and entropy regularization so that explanations are sparse, near-discrete, and compact (Ying et al., 2019).

This baseline matters because later DiGNNExplainer variants preserve its central invariants while changing the graph domain. DGExplainer replaces static neighborhoods by spatio-temporal histories; DRExplainer replaces an undirected computation graph by a directed multi-relational graph; TAGE shifts the target of explanation from a specific downstream predictor to the encoder embedding space; DnX preserves post-hoc explanation but moves it to a distilled surrogate. The inheritance is therefore methodological rather than nominal: local fidelity, mask-based or attribution-based selection, and explanation as a constrained information-preserving substructure remain the recurring motifs (Ying et al., 2019).

3. Dynamic-graph interpretations

In the strict dynamic sense, DiGNNExplainer corresponds most directly to DGExplainer, which models a dynamic graph as a sequence of snapshots

$G = \{G_t\}_{t=1}^T,\quad G_t = (V_t, E_t),$

with adjacency matrices $A_t$ and feature matrices $X_t$. DGExplainer is instantiated on a GCN–GRU architecture: at each time step a multi-layer GCN produces structural embeddings $\hat{X}_t$, and a GRU consumes $\hat{X}_1,\dots,\hat{X}_T$ to generate time-evolving node states. Explanation is then performed by Layer-wise Relevance Propagation, with conservation

$$\sum_{k_1} R_{k_1}^{(l)} = \sum_{k_2} R_{k_2}^{(l+1)} = f(G),$$

and input-level node relevance at time $t$ defined as

$M$0

This yields node-level relevance across time for dynamic link prediction and node regression, rather than a single static subgraph explanation (Xie et al., 2022).

DGExplainer’s significance lies in its explicit treatment of temporal dependence. Static explainers can indicate which nodes are important in the last snapshot, but they do not track which earlier nodes and time steps influenced the final prediction. DGExplainer addresses this by propagating relevance first through the GRU, then through the GCN at every time step. Quantitatively, it achieves the best or comparable fidelity on all six reported datasets, higher sparsity than the baselines, and stability superior to Grad $M$1 Input though slightly inferior to Sensitivity Analysis, making fidelity–stability trade-offs explicit rather than implicit (Xie et al., 2022).

A second dynamic interpretation is DyExplainer, which is not an LRP method but a jointly trained sparse-attention framework for explainable dynamic GNNs. DyExplainer uses a dynamic GNN backbone, a sparse structural attention per snapshot, a temporal attention over a buffer of recent snapshots, and two contrastive priors: structural consistency and temporal continuity. Its total objective is

$M$2

This design turns explanation into part of the predictive pipeline rather than a purely post-hoc diagnostic. On six dynamic link-prediction datasets, DyExplainer outperforms its ROLAND backbone on 4/6 datasets and improves Mean Reciprocal Rank by $M$3 on Bitcoin-Alpha, while also yielding higher Fidelity than GNNExplainer and gradient baselines across reported sparsity levels (Wang et al., 2023).

4. Directed and multi-relational formulations

In the directed sense, DRExplainer provides the clearest concrete realization of what a DiGNNExplainer would be on a directed graph. DRExplainer constructs a directed bipartite network for drug response prediction with node types $M$4 for cell lines and $M$5 for drugs, and four relation types: cell line $M$6 drug sensitive edges, drug $M$7 cell line resistant edges, cell-line similarity edges, and drug-similarity edges. The graph $M$8 is processed by a 2-layer directed relational GCN followed by a DistMult decoder with score

$M$9

so that “directed prediction” becomes the estimation of $\tilde{A} = A_c \odot \sigma(M).$0 and $\tilde{A} = A_c \odot \sigma(M).$1 as distinct triples rather than as a symmetric relation (Shi et al., 2024).

Its explanation mechanism is a direct mask-learning adaptation of GNNExplainer to a directed multi-relational graph. For each relation $\tilde{A} = A_c \odot \sigma(M).$2, DRExplainer learns a continuous mask $\tilde{A} = A_c \odot \sigma(M).$3 and defines

$\tilde{A} = A_c \odot \sigma(M).$4

then optimizes the masked graph so that the original relation prediction is preserved. Because the adjacency is directed, the mask is also directed; no symmetry is imposed. This is precisely the transfer principle from which the paper abstracts a general DiGNNExplainer: operate on the same directed adjacency tensors used by the predictor, preserve edge direction and relation type, and recover a compact explanation by thresholding the learned mask (Shi et al., 2024).

DRExplainer also makes an important methodological contribution by attaching quantitative evaluation to explanation quality. It constructs a benchmark set of ground-truth explanatory edges from cell-line similarity, drug similarity, and response consistency, then evaluates top-$\tilde{A} = A_c \odot \sigma(M).$5 explanation recovery with precision@$\tilde{A} = A_c \odot \sigma(M).$6, recall@$\tilde{A} = A_c \odot \sigma(M).$7, and f1@$\tilde{A} = A_c \odot \sigma(M).$8. On this benchmark, DRExplainer reports precision@$\tilde{A} = A_c \odot \sigma(M).$9 $G = \{G_t\}_{t=1}^T,\quad G_t = (V_t, E_t),$0, recall@$G = \{G_t\}_{t=1}^T,\quad G_t = (V_t, E_t),$1 $G = \{G_t\}_{t=1}^T,\quad G_t = (V_t, E_t),$2, and f1@$G = \{G_t\}_{t=1}^T,\quad G_t = (V_t, E_t),$3 $G = \{G_t\}_{t=1}^T,\quad G_t = (V_t, E_t),$4, versus ExplaiNE’s $G = \{G_t\}_{t=1}^T,\quad G_t = (V_t, E_t),$5, $G = \{G_t\}_{t=1}^T,\quad G_t = (V_t, E_t),$6, and $G = \{G_t\}_{t=1}^T,\quad G_t = (V_t, E_t),$7, respectively. The case studies further show sensitive and resistant predictions supported by similarity-driven directed subgraphs, such as OCIAML2 predicted sensitive to 5-Fluorouracil and OCILY19 predicted resistant to Etoposide (Shi et al., 2024).

5. Task-agnostic and surrogate-based reinterpretations

A broader interpretation treats DiGNNExplainer as an explainer architecture decoupled from any single downstream task. TAGE formalizes this most explicitly. It decomposes explanation into a downstream explainer $G = \{G_t\}_{t=1}^T,\quad G_t = (V_t, E_t),$8, which produces an importance vector over embedding dimensions using gradient-based saliency, and an embedding explainer $G = \{G_t\}_{t=1}^T,\quad G_t = (V_t, E_t),$9, which maps a condition vector $A_t$0 and a graph $A_t$1 to an explanatory subgraph. Training is self-supervised: the explainer maximizes mutual information between masked encoder embeddings of the full graph and the selected subgraph, using conditioned contrastive estimators such as Jensen–Shannon Estimator or InfoNCE, plus edge-size and entropy regularization (Xie et al., 2022).

This changes the unit of explanation from “one model, one task” to “one encoder, many possible tasks.” TAGE is therefore compatible with multitask models and self-supervised encoders whose downstream heads are unknown at explainer-training time. Its empirical significance is clearest on PPI with 121 tasks: the estimated total time to explain all tasks is about $A_t$2 days for GNNExplainer, about $A_t$3 days for PGExplainer, and about $A_t$4 days for TAGE, corresponding to reported speedups of about $A_t$5 and $A_t$6, respectively, while maintaining explanation quality that is as good as or better than task-specific baselines (Xie et al., 2022).

DnX and FastDnX push this decoupling further by replacing the original GNN altogether during explanation. DnX first distills the teacher GNN $A_t$7 into a simple linear SGC surrogate $A_t$8, then explains $A_t$9 rather than $X_t$0. The node-importance mask is obtained by solving a convex quadratic program over the simplex,

$X_t$1

and FastDnX replaces optimization by a closed-form linear decomposition. Theoretical results bound explanation unfaithfulness for the teacher by the surrogate term plus $X_t$2, where $X_t$3 is the distillation error, and the paper reports speedups up to $X_t$4 over GNNExplainer while often outperforming state-of-the-art explainers in accuracy, fidelity, and runtime (Pereira et al., 2023).

6. Counterfactual, causal, and distribution-aware extensions

Several related directions extend the DiGNNExplainer idea beyond fidelity-only subgraph recovery. D4Explainer introduces discrete denoising diffusion on graph structure and combines a distribution-learning term with a counterfactual term,

$X_t$5

so that explanations are not only label-changing or class-supporting but also in-distribution with respect to the training graphs. It is presented as the first unified framework combining counterfactual and model-level explanations. Empirically, it reports the best CF-ACC AUC and Fidelity AUC on $X_t$6 datasets and achieves the lowest or second-lowest MMD to the data distribution across Mutag, BBBP, NCI1, and BA-3Motif, but it also inherits the $X_t$7 cost of its PPGN denoiser and requires choosing the node count $X_t$8 for model-level motifs (Chen et al., 2023).

CIExplainer++ shifts the emphasis from perturbation faithfulness to causal estimands and verbal interpretability. CIExplainer defines feature-level causal effects under the Potential Outcome Framework,

$X_t$9

aggregates them to a node-level score $\hat{X}_t$0, and selects the top-$\hat{X}_t$1 nodes in the computation graph as the explanation subgraph. G2TeXplainer then converts the resulting causal subgraph into motif-aware natural language. On BA-Shapes and Tree-Grid, CIExplainer is reported as best or near-best across GCN, GAT, GIN, and GraphSAGE, while the motif-aware P4 prompt in G2TeXplainer yields the strongest textual scores, including Structure $\hat{X}_t$2, Clarity $\hat{X}_t$3, and Pairwise Win-rate $\hat{X}_t$4 in the reported evaluation (Caldas et al., 17 Jun 2026).

Taken together, these developments clarify the present state of DiGNNExplainer as a field concept rather than a single method. Dynamic variants highlight temporal dependence and snapshot continuity; directed variants preserve edge orientation and relation semantics; task-agnostic and surrogate variants decouple explanation from a fixed head or a complex teacher; diffusion and causal variants add in-distribution counterfactuality or explicit causal-effect semantics. The recurring limitations are equally consistent across the literature: dependence on the backbone architecture, memory and compute costs on large graphs or long sequences, domain-specific ground-truth construction, and the fact that post-hoc explanations remain faithful to what the GNN has learned, not necessarily to the underlying causal mechanism of the domain (Xie et al., 2022, Wang et al., 2023, Shi et al., 2024, Xie et al., 2022, Pereira et al., 2023, Chen et al., 2023, Caldas et al., 17 Jun 2026).