Graph Neural Networks & Node Classification
Architecture-Agnostic Black-Box Explanation for Graph Neural Networks
Generated automatically from the limitations stated in 3 papers (NeurIPS, ICML), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Many current GNN explainers require differentiable access to intermediate node embeddings, meaning they cannot operate under true black-box conditions such as inference-only APIs or proprietary model deployments. Because these methods are tied to specific encoder representations, their explanations fail to generalize across diverse GNN architectures and degrade when underlying embeddings are noisy or poorly calibrated. Practitioners with black-box access to graph classification services are currently blocked from generating post-hoc explanations without access to internal model parameters and activations.
Why it matters
Enables trustworthy post-hoc explanations for closed-source, proprietary, or heterogeneous GNN prediction pipelines where internal embeddings and gradients are strictly unavailable.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Conduct a cross-architecture benchmark evaluating existing perturbation- and search-based explainers strictly through input-output queries (without intermediate embedding access) across multiple GNN backbones (e.g., GCN, GAT, GIN, GraphSAGE), measuring explanation fidelity, sparsity, and query complexity.
- 2
Develop a black-box subgraph attribution method driven purely by output prediction probabilities via discrete sampling or zeroth-order optimization, measuring its stability and explanation accuracy against white-box baselines across diverse node classification tasks.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
The query sample complexity required to obtain high-fidelity subgraph explanations without gradient guidance could be prohibitively high for large-scale graphs, rendering strictly black-box approaches computationally intractable in practice.
Evidence
Each paper's own statement of the limitation, verbatim.
- Generating In-Distribution Proxy Graphs for Explaining Graph Neural NetworksICML 2024
The explainer requires differentiable access to the GNN encoder's node embeddings, so the truly black-box setting emphasized in the motivation is not fully supported in practice
- D4Explainer: In-distribution Explanations of Graph Neural Network via Discrete Denoising DiffusionNeurIPS 2023
Explanations are tied to the specific GNN architecture being explained, limiting generalizability across different GNN models
- Parameterized Explainer for Graph Neural NetworkNeurIPS 2020
Quality of explanations is inherently bounded by the quality of the pre-trained GNN's node embeddings used as input.
Nearest existing work
- Parameterized Explainer for Graph Neural NetworkNeurIPS 2020
- Task-Agnostic Graph ExplanationsNeurIPS 2022
- Distill n’ Explain: explaining graph neural networks using simple surrogatesAISTATS 2023
- GNNExplainer: Generating Explanations for Graph Neural NetworksNeurIPS 2019
- On Explainability of Graph Neural Networks via Subgraph ExplorationsICML 2021
- GNNBoundary: Towards Explaining Graph Neural Networks through the Lens of Decision BoundariesICLR 2024
- Graph Neural Network Explanations are FragileICML 2024
- Adversarial Mask Explainer for Graph Neural NetworksTheWebConference 2024
- GraphFramEx: Towards Systematic Evaluation of Explainability Methods for Graph Neural NetworksLearning on Graphs 2022
- Provably Robust Explainable Graph Neural Networks against Graph Perturbation AttacksICLR 2025
- Towards More Practical Adversarial Attacks on Graph Neural NetworksNeurIPS 2020
- Relevant Walk Search for Explaining Graph Neural NetworksICML 2023
- Redundancy Undermines the Trustworthiness of Self-Interpretable GNNsICML 2025
- Towards Training GNNs Using Explanation Directed Message PassingLearning on Graphs 2022
- PGM-Explainer: Probabilistic Graphical Model Explanations for Graph Neural NetworksNeurIPS 2020