Graph Neural Networks (GNNs) owe their success to a simple yet powerful mechanism: message passing. By iteratively aggregating feature vectors from neighboring nodes along structural edges, GNNs construct representations that blend what an object is (its features) with where it sits in a network (its topology).
However, this blending creates a significant challenge for interpretability. When a GNN makes a prediction—such as classifying a molecule as toxic or flagging a piece of code as vulnerable—traditional explainability tools struggle to explain why. Early interpretability frameworks often fell into a structural-feature explanation gap. They isolated their focus, treating node attributes and topological connectivity as independent variables.
Recent research has shifted toward unified graph explanation, a paradigm that treats features and topology as deeply intertwined, joint contributors to a model's decision-making process.
The Danger of Isolated Explanations
To understand why a unified approach is necessary, consider how an isolated explanation fails. If an interpretability tool focuses exclusively on topology, it might highlight a specific subgraph motif (e.g., a benzene ring in a molecule). However, as Yin et al. and Huang et al. point out, an edge-centric explanation ignores the node-level features that make that structure meaningful. A benzene ring with a chlorine attachment behaves entirely differently than one with a hydroxyl group.
Conversely, focusing solely on node features creates an equally problematic blind spot. As Kasanishi et al. and Duval et al. argue, a node's internal attributes cannot be understood in a vacuum; their impact is dictated by their relational context.
Isolated View (Topology Only) --> [Node] --- (Important Edge) --- [Node] (Missing: What are these nodes?)
Isolated View (Features Only) --> [X: High-Dim Feature Vector] (Missing: How do they interact?)
Unified View --> [Feature X] === (Contextual Edge) === [Feature Y]
When explainers treat these two modalities as separate entities, they create what He et al. term "logical gaps." An explanation might tell you which edges are important and which features are important, but it fails to capture the non-linear interplay between them. For instance, an edge might only be significant because it completes a specific path that allows a critical feature to propagate to the target node.
Mechanism 1: Joint Optimization via Mutual Information
The foundational step toward solving this problem was introduced by Ying et al. through GNNExplainer. Instead of optimizing for structural importance and feature importance sequentially, the framework optimizes for both simultaneously.
GNNExplainer formulates explanation generation as a mutual information maximization problem. Given a trained GNN and a prediction, it learns a compact subgraph mask $M$ and a node feature mask $F$ that maximize the mutual information between the original prediction and the prediction made with the masked input:
$$\max_{M, F} \text{MI}(Y, \hat{Y}(G_c \odot M, X \odot F))$$
By optimizing $M$ and $F$ jointly, the optimization landscape can capture conditional dependencies. An edge mask value is driven not just by the presence of the edge, but by whether the features passing through that edge (regulated by $F$) alter the model's output.
Mechanism 2: Tracking Feature Propagation Through Layers
While masking methods provide a post-hoc look at what matters, they can suffer from unfaithfulness if the perturbations destroy the underlying graph structure in ways the model never encountered during training. To bypass this, Yang et al. developed DEGREE, a decomposition-based framework.
Instead of masking the input, DEGREE tracks how specific input components flow through the message-passing layers. By decomposing the mathematical operations of the GNN layer-by-layer, it maps exactly how a node's internal attributes are transformed as they travel across structural connections.
This approach provides an exact accounting of feature-topology integration: it reveals not just that a neighbor is important, but precisely which of that neighbor's features traveled along the connecting edge to influence the final prediction.
Mechanism 3: Structure-Aware Game Theory
Another prominent angle leverages cooperative game theory to distribute credit fairly among features and structures. A common approach in machine learning is the Shapley value, which calculates the marginal contribution of a feature across all possible coalitions. However, standard Shapley values assume features can be arbitrarily removed or combined—an assumption that breaks down when data is bound by a graph topology.
To address this, Duval et al. introduced GraphSVX, which adapts Shapley values specifically for GNNs by defining coalitions over a joint space of nodes and features. Similarly, Zhang et al. proposed GStarX, which utilizes the structure-aware Hamiache-Navarro (HN) value. By embedding the graph's adjacency matrix directly into the game-theoretic framework, these methods ensure that the computed importance of a node feature is intrinsically weighted by its structural connectivity.
Scaling Up: Parameterized and Inductive Explanations
Early unified explainers required running a costly optimization process from scratch for every single node or graph prediction. To make these methods practical for real-world deployment, researchers shifted toward parameterized, generative models.
Luo et al. and Cheng et al. pioneered this transition with PGExplainer. Instead of optimizing masks for an individual instance, they trained a separate neural network to predict the importance of graph components across the entire dataset.
Instance-based (Slow): [Target Instance] --> [Optimization Loop] --> [Explanation]
Parameterized (Fast): [Target Instance] --> [Trained Explainer Network] --> [Explanation]
This parameterization allows the explainer to learn universal structural-feature patterns. As a result, it enables inductive explanation: the ability to instantly generate unified explanations for entirely new, unseen graphs without retraining, achieving massive speedups over instance-level optimization.
Real-World Impact: Chemistry and Cybersecurity
The necessity of unifying feature and structural attributions becomes obvious when applying GNNs to high-stakes domains:
- Molecular Analysis: In chemistry, predictions rely on both atoms (nodes) and chemical bonds (edges). As noted by Akita et al. and Henderson et al., an explanation that only highlights atoms without explaining the bonds connecting them—or vice versa—is chemically nonsensical. Unified explainers allow medicinal chemists to see the exact functional groups and bond regularizations driving property predictions.
- Cybersecurity: When analyzing code vulnerability graphs, a node might represent a variable, and an edge might represent data flow. He et al. demonstrated that unified frameworks prevent false positives by ensuring that a security analyst sees both the structural execution path and the specific variable attributes simultaneously.
By moving away from isolated, single-component explanations and embracing the joint interplay of structure and features, unified graph explanation frameworks provide the fidelity required to truly trust GNNs in production.
Key Foundational Work
GNNExplainer: Generating Explanations for Graph Neural Networks (2019) — contributed 1 concept elements to this direction.
BayesGrad: Explaining Predictions of Graph Convolutional Networks (2018) — contributed 1 concept elements to this direction.