FSX: Message Flow Sensitivity Enhanced Structural Explainer for Graph Neural Networks

Despite the widespread success of Graph Neural Networks (GNNs), understanding the reasons behind their specific predictions remains challenging. Existing explainability methods face a trade-off that gradient-based approaches are computationally efficient but often ignore structural interactions, while game-theoretic techniques capture interactions at the cost of high computational overhead and potential deviation from the model’s true reasoning path. To address this gap, we propose FSX (Message Flow Sensitivity Enhanced Structural Explainer), a novel hybrid framework that synergistically combines the internal message flows of the model with a cooperative game approach applied to the external graph data. FSX first identifies critical message flows via a novel flow-sensitivity analysis: during a single forward pass, it simulates localized node perturbations and measures the resulting changes in message flow intensities. These sensitivity-ranked flows are then projected onto the input graph to define compact, semantically meaningful subgraphs. Within each subgraph, a flow-aware cooperative game is conducted, where node contributions are evaluated fairly through a Shapley-like value that incorporates both node-feature importance and their roles in sustaining or destabilizing the identified critical flows. Extensive evaluation across multiple datasets and GNN architectures demonstrates that FSX achieves superior explanation fidelity with significantly reduced runtime, while providing unprecedented insights into the structural logic underlying model predictions–specifically, how important sub-structures exert influence by governing the stability of key internal computational pathways.

💡 Research Summary

The paper introduces FSX (Message Flow Sensitivity Enhanced Structural Explainer), a hybrid framework that unifies internal message‑passing dynamics of Graph Neural Networks (GNNs) with external graph‑structure analysis to produce faithful yet efficient explanations. Existing GNN explainers fall into two camps: gradient‑ or perturbation‑based black‑box methods that are fast but ignore the model’s internal reasoning, and game‑theoretic or flow‑based approaches that capture interactions but are computationally prohibitive. FSX bridges this gap by first performing a flow‑sensitivity analysis in a single forward pass. For each layer‑edge triplet (u, v, l), the method temporarily weakens the message from node u to node v by a damping factor γ, recomputes the model’s output, and records the absolute change in the target logit. This change, S(u,v,l), quantifies the causal contribution of that specific message flow to the prediction. By iterating over all triplets, FSX builds a fine‑grained sensitivity map without requiring gradients or multiple backward passes, thus keeping the computational overhead low.

The second stage translates the sensitivity map into an interpretable subgraph. The top‑K most sensitive flows are selected, and the nodes and edges involved form a compact subgraph G_key. This subgraph is guaranteed to contain the structural elements that actually sustain the most influential internal flows. Within G_key, FSX defines a flow‑aware cooperative game: the player set is the nodes of G_key, and the value function ν(S) measures how well a coalition S preserves the identified critical flows, weighting each flow by its sensitivity score and by whether the coalition contains the necessary endpoints. This design ensures that the Shapley‑like attribution reflects both feature importance and the structural role of a node in maintaining the stability of key information pathways.

Because computing exact Shapley values is exponential, the authors adopt a Monte‑Carlo Markov Chain sampling scheme combined with a greedy approximation to estimate a weighted Shapley value for each node. Crucially, the game is confined to G_key, dramatically reducing the combinatorial space compared with whole‑graph approaches such as GraphEXT or FlowX. Empirical evaluation on standard node‑classification datasets (Cora, Citeseer, PubMed) and graph‑classification benchmarks (MUTAG, PROTEINS) across several GNN architectures (GCN, GAT, GIN) shows that FSX achieves higher explanation fidelity (measured by the drop in prediction confidence after removing the explained subgraph) while cutting runtime by 5–10× relative to prior state‑of‑the‑art methods. The resulting explanations are also more concise, typically involving fewer than 20 % of the original nodes, and they clearly distinguish nodes that are essential for preserving critical flows from those that merely contribute peripheral information.

The paper acknowledges limitations: the exhaustive enumeration of all (u, v, l) triplets can become costly for very deep networks or massive graphs, the choice of the damping factor γ influences sensitivity scores and may require tuning, and the Shapley approximation introduces a trade‑off between accuracy and speed. The authors suggest future work on flow‑sampling heuristics, adaptive γ selection, and distributed implementations to scale FSX to industrial‑size graphs.

In summary, FSX presents a novel, principled approach that tightly couples a GNN’s internal computational pathways with external structural explanations, delivering a compelling balance of interpretability, fidelity, and efficiency that advances the state of explainable graph learning.