Joint Structural and Feature Explanations for Graph Neural Networks

This direction is merged into another direction: model interpretability (GNN interpretability), unified graph explanation. Its last version stays readable and citable here.

Version 1 of 130 papers38 conceptsfrozen 2026-10-04

Graph Neural Networks (GNNs) are powerful because they process two distinct types of information simultaneously: the attributes of individual entities (node features) and the relationships between them (graph topology). However, this dual nature makes them notoriously difficult to interpret. To explain a GNN's prediction, one must determine whether the model was triggered by a specific node's properties, a particular connection, or a complex motif formed by both.

The research direction in GNN interpretability has evolved from identifying isolated important edges to creating unified frameworks that capture the interplay between structure and features.

The Core Mechanism: Masking and Mutual Information

The foundational approach to GNN explanation, pioneered by GNNExplainer by Rex Ying et al., treats the problem as an optimization task. To explain a prediction, the method searches for a "mask"—a small subset of edges and node features—that preserves the model's original output.

The intuition is rooted in Mutual Information (MI). By maximizing the MI between the original prediction and the prediction made on the masked graph, the explainer identifies the "minimal sufficient" subgraph and feature set. This approach treats topology and features as joint contributors, recognizing that an edge might only be significant if the nodes it connects possess specific attributes.

From Instance-Specific to Parameterized Explanations

Early methods like GNNExplainer were "one-off" optimizations: to explain a single node's classification, you had to run a new optimization process from scratch. This is computationally expensive and fails to generalize to unseen data.

PGExplainer, developed by Dongsheng Luo, Wei Cheng, and colleagues, shifted this paradigm toward parameterization. Instead of optimizing a mask for each instance, they trained a separate neural network (the "explainer") to predict which edges are important across the entire dataset. This "train once, explain everywhere" philosophy allows for inductive explanations, where the model can instantly explain predictions on new graphs it has never seen before, achieving massive speedups.

The Duality of Structure and Features

A recurring tension in this field is the balance between structural and feature-based explanations.

  • Feature-Centric: Methods like GraphLIME by Qiang Huang et al. adapt local linear models (like LIME) to the graph domain, focusing on which node attributes drive a decision within a local neighborhood.
  • Structure-Centric: Conversely, StGraphLIME by Ayato Toyokuni and Makoto Yamada emphasizes that in many domains, the arrangement of nodes is more informative than the nodes themselves.

However, Steve Azzolin and colleagues point out in their work on Dual-Channel GNNs (DC-GNNs) that purely structural explanations fail when a task depends on global properties or feature-counting. For example, if a GNN identifies a "toxic" molecule based on the total number of oxygen atoms rather than a specific ring structure, a purely topological explanation will be unfaithful to the model's logic. This has led to hybrid approaches like SAME by Ziyuan Ye et al., which uses Shapley-based values to account for the topological arrangement of importance.

Explaining through Change: Counterfactuals

A more recent frontier is counterfactual explanation, which asks: "What is the smallest change I can make to this graph to flip the model's prediction?"

Methods like COMBINEX by Fabrizio Silvestri et al. jointly perturb both edges and node features to find these minimal changes. Similarly, Chirag Chhablani and colleagues utilize game-theoretic concepts like the Banzhaf value to identify robust counterfactuals. This approach is intuitive for humans: if removing a specific friendship link changes a "high-risk" social network classification to "low-risk," that link is clearly the most actionable part of the explanation.

Generalization and Pre-training

The field is currently moving toward universal interpretability. $\pi$-GNN, introduced by Jun Yin and Chaozhuo Li, suggests that interpretability can be pre-trained. By learning universal structural motifs from synthetic data, these models can provide explanations across entirely different datasets, suggesting that the "logic" of graph connectivity may be more transferable than previously thought.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_joint_structural_and_feature_explanation_05304175,
  title = {Joint Structural and Feature Explanations for Graph Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/joint-structural-and-feature-explanations-for-graph-neural-networks/rdr_053041757b60468c9b80697df93dc53a},
  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_joint_structural_and_feature_explanation_05304175_v1,
  title = {Joint Structural and Feature Explanations for Graph Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/joint-structural-and-feature-explanations-for-graph-neural-networks/rdr_053041757b60468c9b80697df93dc53a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}