Self-Denoising in SI-GNN Explanations

Updated 5 July 2026

Self-Denoising (SD) is a post-processing technique for SI-GNNs that leverages self-inconsistency in edge masks to recalibrate and enhance explanation stability.
It computes an instability vector from dual-pass edge scores and suppresses context-driven edges, ensuring only stable signals contribute to the final explanation.
The method is model-agnostic and adds only 4–6% computational overhead while consistently improving explanation quality and prediction fidelity.

Self-Denoising (SD) is a post-processing strategy for explanations produced by Self-Interpretable Graph Neural Networks (SI-GNNs). In this setting, the method exploits self-inconsistency: when an SI-GNN is reapplied to its own explanatory graph subset, the resulting explanation can differ from the original one even though the model parameters have not changed. SD treats that instability as an intrinsic denoising signal, using one additional forward pass to suppress context-driven edges that do not provide stable evidence for the prediction. It is model-agnostic, training-free, and, in experiments, adds only about 4–6% computational overhead while consistently improving explanation quality across representative SI-GNN frameworks, backbone architectures, and benchmark datasets (Tai et al., 8 May 2026).

1. Problem setting in self-interpretable GNNs

In a Self-Interpretable Graph Neural Network, the model consists of an explainer, an encoder, and a classifier. It first produces a soft mask $M^{(1)}$ , yielding an explanatory subgraph

$G_s^{(1)} = G \odot M^{(1)},$

and then predicts from $G_s^{(1)}$ . If the same explainer is reapplied to $G_s^{(1)}$ , it produces a new mask $M^{(2)}$ for the same original graph $G$ . Empirically, $M^{(2)}$ often differs from $M^{(1)}$ , which is the phenomenon termed self-inconsistency (Tai et al., 8 May 2026).

This observation matters because SI-GNN explanations are intended to identify the stable evidence supporting a prediction. The motivating premise of SD is that self-inconsistent edges do not provide stable evidence for the model’s prediction. Rather than discarding the inconsistency as an implementation artifact, the method uses it to recalibrate the original explanation.

2. Re-explanation-induced context perturbation

The direct cause of score variation is re-explanation-induced context perturbation. Masking changes the weights on edges, which alters local message-passing contexts. Since edge scores are computed from node embeddings aggregated over neighborhoods, re-masking perturbs those embeddings and thus perturbs scores (Tai et al., 8 May 2026).

This mechanism explains why self-inconsistency can arise even when the explainer, encoder, and classifier parameters remain fixed. The second explanation is not produced under the same effective neighborhood structure as the first one. The perturbation is therefore endogenous to the explanatory workflow itself: explanation changes the graph context, and the changed context changes subsequent explanation scores.

3. Latent signal assignment hypothesis

To explain why only a subset of edges vary strongly under re-explanation, the method introduces a latent signal assignment hypothesis. Under this view, edges occupy three latent states.

Edge state	Score range	Characterization
Positive-signal edges ( $E_p$ )	$[x_+,1]$	Stable high score; intrinsic model signal independent of neighborhood context
Negative-signal edges ( $G_s^{(1)} = G \odot M^{(1)},$ 0)	$G_s^{(1)} = G \odot M^{(1)},$ 1	Stable low score; intrinsic negative signal
Context-driven edges ( $G_s^{(1)} = G \odot M^{(1)},$ 2)	$G_s^{(1)} = G \odot M^{(1)},$ 3	No intrinsic signal; score depends heavily on the current masked neighborhood

Positive-signal edges typically coincide with truly important edges, while negative-signal edges remain stably unimportant. Context-driven edges lack intrinsic support and are the ones that “move around” under re-explanation. The paper also analyzes how conciseness regularization affects latent signal assignment (Tai et al., 8 May 2026).

This hypothesis gives SD its central interpretive claim: instability is not uniformly distributed across the explanation. It is concentrated on edges whose scores are dominated by context rather than by intrinsic model signals. A plausible implication is that self-inconsistency can be used as a selective filter rather than as a global confidence penalty.

4. Mathematical formulation and algorithm

SD compares first-pass and second-pass edge scores and shrinks edges whose scores change substantially. Let

$G_s^{(1)} = G \odot M^{(1)},$ 4

be the first-pass edge scores and

$G_s^{(1)} = G \odot M^{(1)},$ 5

the second-pass scores on $G_s^{(1)} = G \odot M^{(1)},$ 6. Define the instability vector

$G_s^{(1)} = G \odot M^{(1)},$ 7

The calibrated score is

$G_s^{(1)} = G \odot M^{(1)},$ 8

Here $G_s^{(1)} = G \odot M^{(1)},$ 9 is the denoising strength. The operational intuition is direct: edges that “flip around” under re-explanation, and therefore have large $G_s^{(1)}$ 0, are multiplicatively shrunk (Tai et al., 8 May 2026).

For a single graph, the post-processing procedure is:

Run the explainer to obtain the soft mask $G_s^{(1)}$ 1.
Construct $G_s^{(1)}$ 2.
Re-run the explainer on $G_s^{(1)}$ 3 to obtain $G_s^{(1)}$ 4.
Compute instability $G_s^{(1)}$ 5.
Calibrate with

$G_s^{(1)}$ 6

Return $G_s^{(1)}$ 7 as the denoised explanation.

No retraining of model parameters is needed. Only the explainer component is invoked twice, so the method requires exactly one extra forward pass.

5. Denoising strength, classifier adaptation, and implementation

The choice of $G_s^{(1)}$ 8 controls the trade-off between insufficient suppression and over-suppression. If $G_s^{(1)}$ 9 is too small, context-driven edges may remain too large; if $G_s^{(1)}$ 0 is too big, even genuine signals can be over-suppressed. Propositions 4.1–4.2 provide analytic bounds on when SD corrects pairwise mis-rankings and on limiting the prediction shift (Tai et al., 8 May 2026).

In practice, the recommended procedure is to freeze the explainer and encoder, fine-tune only the classifier for 10 epochs on a validation split of SD-calibrated masks, and select the $G_s^{(1)}$ 1 that maximizes validation accuracy. This lightweight adaptation is referred to as SD*. Its role is to mitigate the distribution shift introduced by SD.

The computational profile is intentionally narrow. SD requires exactly one extra forward pass through the explainer and encoder each time an explanation is denoised, and across all tested architectures this adds only about 4–6% wall-clock overhead compared to a single explanation. No gradient calculations or retraining of the explainer or encoder are performed; only inference is added, plus optional classifier fine-tuning for $G_s^{(1)}$ 2 selection. In the reported implementation, edge masks are trained with Gumbel-Sigmoid, while inference uses a deterministic $G_s^{(1)}$ 3 mask; SD then post-multiplies that mask by $G_s^{(1)}$ 4 (Tai et al., 8 May 2026).

A common misconception is to read SD as a modified explainer-training method. In the formulation above, it is instead a post-processing method applied to any pretrained SI-GNN.

6. Empirical behavior, complementarity, and terminological scope

The empirical study covers four representative SI-GNN paradigms—GAT, CAL, SMGNN, and GSAT—three backbone architectures—GIN, GraphSAGE, and GatedGCN—and four benchmark datasets: BA-2MOTIFS, 3MR, BENZENE, and MUTAGENICITY. The reported metrics are AUC, Fidelity $G_s^{(1)}$ 5, ACC, and SPA, where AUC measures plausibility against human-annotated ground truth, Fidelity $G_s^{(1)}$ 6 measures prediction stability on $G_s^{(1)}$ 7, ACC measures the model’s accuracy on $G_s^{(1)}$ 8 as informativeness, and SPA is the average mask weight as conciseness. On the GIN backbone, AUC improved in all 16 model–dataset combinations, by up to +3.9%; SPA dropped consistently, indicating more concise explanations; Fidelity $G_s^{(1)}$ 9 sometimes slightly worsened without classifier adaptation because of distribution shift, but after SD* it was as good or better; and ACC was largely unchanged or slightly improved, rising by up to +8% after SD*. These gains also generalize to GraphSAGE and GatedGCN (Tai et al., 8 May 2026).

SD is also complementary to Explanation Ensemble (EE). The combination SD+EE gives strictly higher AUC and ACC than EE alone, while adding only 4–6% extra cost versus EE’s ~400% cost. This positions SD as a low-overhead calibration stage rather than as an alternative to ensemble-based explanation improvement.

The expression “Self-Denoising” is used in other literatures with a different meaning. In image and medical-imaging research, it denotes self-supervised denoising from noisy observations, including the joint blind denoising and noise estimation framework of Ollion et al. (Ollion et al., 2021), the compressive autoencoder framework for massive noisy images (Wang et al., 2021), the DTI method SDnDTI (Tian et al., 2021), and the score-based MRI framework Corruption2Self (Tu et al., 8 May 2025). In those cases, denoising refers to restoring signals corrupted by measurement noise. In the SI-GNN setting, by contrast, SD denoises explanations rather than images: it suppresses context-driven explanatory edges by exploiting the model’s own self-inconsistency (Tai et al., 8 May 2026).