Explainable Graph Neural Networks

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2847 papers943 conceptsfrozen 2026-10-04

Graph Neural Networks (GNNs) have become the standard for modeling relational data, from molecular structures to financial transaction webs. However, their architecture—which aggregates information across complex, multi-hop neighborhoods—renders them "black boxes." In high-stakes domains like medicine and finance, a prediction without a justification is often unusable.

The research direction of Explainable Graph Neural Networks (XGNN) seeks to bridge this gap. Rather than just providing a label (e.g., "this molecule is toxic" or "this loan is high-risk"), these methods identify the specific subgraphs, node features, or "what-if" scenarios that drove the model's decision.

The Search for Influential Subgraphs

The most intuitive way to explain a GNN's prediction is to identify a "factual" explanation: the minimal subset of the input graph that preserves the original prediction. If a model classifies a molecule as a mutagen, the explanation should ideally be the specific functional group responsible for that property.

To find these subgraphs, researchers have moved beyond simple gradient-based methods. Wanyu Lin, Hao Lan, and Baochun Li introduced Gem, which uses Granger causality to train a generative model that identifies causal subgraphs. By treating the explanation task as a causal discovery problem, Gem can provide insights significantly faster than previous optimization-based approaches.

Similarly, Jianlong Zhou and colleagues developed GANExplainer, which uses Generative Adversarial Networks to "pit" a subgraph generator against a discriminator. This process forces the generator to produce compact, highly accurate subgraphs that the target GNN views as the "essence" of the original input. For large-scale applications, sGNN-LRP reframes the explanation process as a message-passing problem, allowing for subgraph attribution in linear rather than exponential time.

Counterfactuals and What-If Reasoning

A different facet of interpretability is counterfactual explanation: identifying the smallest change to a graph that would flip the model's prediction. This is particularly valuable in finance and law, where stakeholders need to know how to change an outcome (e.g., "What must change in this supply chain to lower the risk score?").

Jing Ma proposed CLEAR, a generative framework that uses Variational Autoencoders (VAEs) to create these "what-if" scenarios. Unlike factual explanations that look for existing evidence, CLEAR generates synthetic graph perturbations that respect the data's underlying causal distribution. More recently, Fabrizio Silvestri and Flavio Giorgi introduced COMBINEX, which jointly optimizes perturbations to both the graph's structure (edges) and its node features, providing a more holistic view of model sensitivity.

Interpretable by Design

While post-hoc explainers try to demystify existing models, a parallel effort focuses on building GNNs that are inherently transparent. These "self-explainable" models avoid the "black box" problem by making their internal reasoning steps human-readable.

One approach is to use Concept Bottleneck Models. Pietro Barbiero, Gabriele Ciravegna, and Giuseppe Marra developed Relational Concept Bottleneck Models (R-CBMs), which force the GNN to first identify high-level, human-understandable concepts before making a final prediction. Similarly, Jonas Jürß introduced HELP, which uses hierarchical latent pooling to mimic human reasoning—building from small motifs to high-level structural concepts.

In the context of link prediction, Huaisheng Zhu and Suhang Wang developed ILP-GNN, which identifies "bridge neighbors" to justify why two nodes should be connected, providing built-in structural evidence for every predicted link.

Trust and Reliability in High-Stakes Domains

The push for explainability is driven by the need for Trustworthy Graph Learning (TwGL). As noted by Bingzhe Wu and Ran He, explainability is a prerequisite for reliability and privacy.

In healthcare, Pietro Barbiero and Pietro Lio have applied these concepts to "medical digital twins," where GNNs forecast disease trajectories. Here, transparency is not just a feature but a safety requirement, ensuring clinicians can verify the biological logic behind a forecast. In finance, researchers like Shuo Yang and Jun Zhou at Ant Financial use GNNs to solve data deficiency in SME lending, where the ability to explain a credit flag is essential for regulatory compliance and auditing.

Ultimately, the goal of this research direction is to move GNNs from powerful but opaque statistical engines to collaborative tools that provide both a prediction and the "why" behind it.

Papers in this version

Showing 60 of 847 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 784 papers
  2. Version 1 · 2026-10-04 · 847 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_model_interpretability_gnn_interpretabil_269d399d,
  title = {model interpretability (GNN interpretability)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/model-interpretability-gnn-interpretability/rdr_269d399d32684c8f8ff5929824fd195c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_explainable_graph_neural_networks_269d399d_v1,
  title = {Explainable Graph Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/explainable-graph-neural-networks/rdr_269d399d32684c8f8ff5929824fd195c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}