Explainable Graph Neural Network

A model-agnostic, post-hoc explanation method that identifies a compact subgraph and a small subset of node features most influential for a GNN's prediction. For a given node, it learns a soft mask over edges and features by maximizing mutual information between the original prediction and the prediction on the masked graph. In cellular topology, this reveals which neighboring base stations and which specific features (e.g., transmit power, queue length) drove a congestion classification.

A feature attribution method adapted to graph inputs that assigns importance scores to input features by accumulating gradients along a path from a neutral baseline to the actual input. For a GNN predicting interference, it quantifies how much each edge feature (e.g., path loss) or node feature contributed to the output relative to a zero-information baseline. Satisfies the completeness axiom, ensuring attributions sum to the prediction difference.

Generates minimal perturbations to the input graph that flip the GNN's prediction, answering 'what would need to change for a different outcome?' For a base station classified as 'overloaded', a counterfactual might identify that removing a specific interference edge or reducing the load on one neighboring node by 15% would change the classification to 'normal'. These are actionable insights for network operators.

Applies cooperative game theory to assign a Shapley value to each node or edge, representing its marginal contribution to the prediction across all possible coalitions. GraphSVX and similar methods approximate these values efficiently. In anomaly detection on a cellular topology graph, this pinpoints the exact set of base stations whose collective behavior constitutes the anomalous pattern, providing a mathematically rigorous attribution.

Leverages the learned attention coefficients in Graph Attention Networks (GATs) as a built-in explanation mechanism. Higher attention weights between two base stations indicate the model deemed that relationship more relevant for the task. While intuitive, caution is warranted: attention weights do not always correlate with feature importance and can be misleading if treated as definitive causal explanations without further validation.

Moves beyond low-level edge and node attributions to test if a GNN has learned high-level, human-understandable concepts. Techniques probe the latent space to see if specific neurons or directions correspond to concepts like 'high-interference cluster' or 'dense urban topology'. This allows verification that the model's internal reasoning aligns with domain knowledge about radio network behavior.