The GECo algorithm for Graph Neural Networks Explanation

Graph Neural Networks (GNNs) are powerful models that can manage complex data sources and their interconnection links. One of GNNs’ main drawbacks is their lack of interpretability, which limits their application in sensitive fields. In this paper, we introduce a new methodology involving graph communities to address the interpretability of graph classification problems. The proposed method, called GECo, exploits the idea that if a community is a subset of graph nodes densely connected, this property should play a role in graph classification. This is reasonable, especially if we consider the message-passing mechanism, which is the basic mechanism of GNNs. GECo analyzes the contribution to the classification result of the communities in the graph, building a mask that highlights graph-relevant structures. GECo is tested for Graph Convolutional Networks on six artificial and four real-world graph datasets and is compared to the main explainability methods such as PGMExplainer, PGExplainer, GNNExplainer, and SubgraphX using four different metrics. The obtained results outperform the other methods for artificial graph datasets and most real-world datasets.

💡 Research Summary

The paper introduces GECo (Graph Explainability by Communities), a novel method for interpreting the predictions of Graph Neural Networks (GNNs) in graph‑classification tasks. The core intuition is that densely connected substructures—communities—play a decisive role in the message‑passing process of GNNs, and therefore their presence should strongly influence the final classification. GECo operationalizes this intuition through a five‑step pipeline: (1) the trained GNN is used to obtain the predicted class and probability for the whole input graph; (2) the graph is partitioned into communities using the Louvain (Blondel) modularity‑optimization algorithm; (3) each community is isolated as a subgraph and fed back into the same GNN, recording the probability assigned to the original predicted class; (4) an average (or median) of these community‑level probabilities is computed to define a threshold τ; (5) all communities whose probability exceeds τ are merged, and the union of their nodes constitutes the final explanation mask. This mask highlights the most influential substructures while discarding irrelevant parts of the graph.

GECo belongs to the perturbation‑based, instance‑level family of GNN explainers, but it differs from prior work (e.g., GNNExplainer, PGExplainer, SubgraphX, PGMExplainer) by operating at the community level rather than at the individual node or edge level. This intermediate granularity yields two practical benefits: (i) it reduces the number of forward passes required for explanation (only one pass per community instead of per node/edge), which dramatically lowers computational cost on large graphs; (ii) it aligns with the hierarchical nature of many real‑world graphs where functional modules are naturally clustered.

The authors evaluate GECo on six synthetic datasets and four real‑world molecular datasets. Synthetic datasets are built from Erdős‑Rényi (ER) and Barabási‑Albert (BA) graphs, each augmented with a known motif (house, cycle‑5, cycle‑6, wheel, or grid). Because the ground‑truth explanatory subgraph is known, metrics such as Fidelity⁺/⁻ (how predictions change when the mask is removed or retained) and Intersection‑over‑Union (IoU) can be directly measured. GECo consistently outperforms the baseline explainers on these metrics, especially in cases where the discriminative motif forms a clear community.

Real‑world experiments involve molecular graphs where functional groups (e.g., benzene rings, fluoride‑carbonyl groups) serve as ground‑truth explanations. GECo achieves higher recall and precision than competing methods, notably attaining a recall of 0.92 on the Benzene dataset, indicating that the community‑based mask successfully captures chemically relevant substructures.

Despite its strengths, the paper acknowledges several limitations. Community detection is sensitive to graph topology; overlapping or poorly defined communities may cause important nodes to be omitted from the final mask. The use of a simple mean‑based threshold τ may not be optimal across diverse datasets, suggesting a need for adaptive or learned thresholds. Moreover, the experimental validation is limited to Graph Convolutional Networks (GCNs); extending GECo to attention‑based GNNs (GAT), sampling‑based models (GraphSAGE), or heterogeneous GNNs remains an open question.

Future work directions proposed include: (1) incorporating multi‑scale or overlapping community detection to capture hierarchical structures; (2) learning τ jointly with the explanation process, perhaps via a small auxiliary network; (3) benchmarking GECo on large‑scale non‑molecular graphs such as social networks, citation networks, or knowledge graphs; and (4) integrating the method with model‑agnostic explanation frameworks to broaden its applicability.

In summary, GECo offers a conceptually simple yet effective approach to GNN interpretability by leveraging community structure. It delivers competitive performance with lower computational overhead, and its design opens several promising avenues for further research and practical deployment in domains where understanding graph‑based decisions is critical.