Keywords: Graph Neural Networks, Interpretability, Explainability
Abstract: Graph Neural Networks (GNNs) are powerful models that manage complex data sources and their interconnection links. One of GNNs' main drawbacks is their lack of interpretability, which limits their applicability in sensitive cases. In this paper, we introduce a new methodology involving graph communities to address the interpretability of graph classification problems. The proposal, called GECo (Graph Explanation by COmmunities), exploits the idea that a community, i.e., a subset of graph nodes densely connected, should play a crucial role in graph classification. This assumption is reasonable considering the message-passing mechanism, the core of GNNs. GECo analyzes the contribution to the classification result of the community graphs, building a mask that highlights graph-relevant structures. It first uses the trained GNN one wants to explain to classify the entire graph. Then, it detects the different communities; for each community, a smaller subgraph, including the community nodes’ is created, and the trained GNN is run to see how likely the subgraph alone supports the predicted class. After evaluating all the subgraph communities, an average probability is calculated and set as a threshold. Finally, any subgraph community with a probability value higher than the threshold is assessed as necessary for the model's decision. The collection of these key communities is the basis for the final explanation since they allow the highlighting of the most relevant parts of the graph leading to the classification. GECo has been tested on GNN employing Graph Convolutional Networks layers, using six artificial and four real-world graph datasets. The six synthetic datasets were generated by adding some artificial motifs (e.g., house, cycle, etc.) to Erdos-Renyi and Barabasi-Albert graphs. The real-world datasets contain molecule structures. Both categories of datasets are adopted in the experimental part of the state-of-the-art proposals for graph explainability. GECo has been compared with a random baseline explainer and four state-of-the-art approaches: PGExplainer, PGMExplainer, GNNExplainer, and SubgraphX. We chose these methods for their different strengths, specifically PGExplainer for its efficiency and generalization capability through a learned explanation model, PGMExplainer for its probabilistic approach based on causal graphs, GNNExplainer for its detailed subgraph and feature-level explanations, and SubgraphX for its theoretically grounded subgraph selection by Shapley values. These choices ensure a comprehensive evaluation of our approach against a wide range of robust techniques. We assessed GECo's performance using four evaluation criteria that leverage predicted and ground-truth explanations and use user-controlled parameters, such as the probability distribution obtained by the GNN. The results obtained by GECo consistently outperform state-of-the-art techniques across multiple metrics for synthetic and most real-world datasets. In addition, GECo is significantly faster than its competitors in terms of computational efficiency, making it an ideal solution for large-scale data analysis and practical applications. These strengths solidify GECo’s role in generating accurate, efficient, and interpretable explanations in graph-based classification tasks.
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 11098
Loading