GraphMask

GraphMask introduces a learned, continuous mask over the messages passed between nodes in each GNN layer. Unlike hard pruning, this soft masking uses a sigmoid gating function parameterized by a neural network, allowing the model to be trained end-to-end with gradient descent. The mask values are bounded between 0 and 1, where a value of 0 completely suppresses a message, and 1 allows it to pass unchanged. This differentiability is critical for optimizing the mask to balance fidelity and sparsity.

GraphMask computes a separate mask for each layer of the GNN, providing a granular view of how information flows through the network's depth. This allows practitioners to see not just which input edges matter, but which intermediate representations are critical. For example, in a 3-layer GCN, GraphMask reveals that layer 1 might focus on local atomic bonds, while layer 3 aggregates higher-order molecular substructures. This layer-wise decomposition is essential for debugging deep graph architectures.

The training objective is grounded in the Information Bottleneck principle. GraphMask is trained to maximize the mutual information between the masked graph representation and the original model's prediction, while simultaneously penalizing the number of non-zero masks. This creates a pareto-optimal trade-off where the explanation subgraph retains only the most predictive structural information. The resulting mask is a faithful proxy for the original model's reasoning, not just a correlated artifact.

To avoid costly per-example optimization, GraphMask uses an amortized inference model. A separate neural network, the mask predictor, is trained to predict the edge masks for any input graph in a single forward pass. This contrasts with methods like GNNExplainer that require iterative optimization for each new example. Once trained, GraphMask provides near-instantaneous explanations, making it suitable for production debugging on large-scale graph datasets.

A key innovation is the use of a learned baseline vector to replace masked messages, rather than simply zeroing them out. If a message is fully masked, it is substituted with this learned, layer-specific representation. This prevents the introduction of an out-of-distribution 'zero' signal that could artificially distort the GNN's activations. The baseline is optimized jointly with the masks to represent a truly neutral, uninformative message, ensuring the explanation reflects genuine irrelevance.

GraphMask operates as a post-hoc, plug-in module that does not require modifying the original GNN's architecture or retraining its weights. It intercepts the messages passed between the original model's layers. This makes it compatible with a wide range of message-passing frameworks, including GCNs, GATs, and RGCNs. The original model is treated as a frozen black box, preserving the exact predictive behavior that the user needs to explain.

A model-agnostic framework that identifies a compact subgraph structure and a small subset of node features most influential for a GNN's prediction. Unlike GraphMask's layer-wise masking approach, GNNExplainer directly optimizes a continuous mask on the input adjacency matrix and feature space to maximize mutual information with the model's output. It provides a single, global explanation subgraph rather than analyzing message passing across individual layers.

An explainability method that uses Monte Carlo Tree Search (MCTS) to efficiently explore and identify the most critical subgraph structures for GNN predictions. SubgraphX evaluates subgraph importance using Shapley values as the scoring function, guiding the MCTS to prune unpromising regions of the search space. This contrasts with GraphMask's gradient-based mask learning by employing a discrete, reward-driven exploration strategy that directly measures cooperative game-theoretic contributions.

A quantitative evaluation score that measures how accurately an explanation subgraph reflects the true reasoning process of the GNN. Faithfulness is typically assessed by the drop in prediction performance when the identified important nodes or edges are removed from the original graph. A faithful explanation should cause a significant drop, indicating it captured genuinely relevant structures. This metric is a primary benchmark for evaluating GraphMask's sparse masks against ground-truth explanations.

A self-explainable GNN framework where a generator module extracts a concise, causal subgraph (the rationale) and a predictor module makes a decision based solely on that rationale. This is an inherently interpretable architecture, unlike GraphMask's post-hoc analysis. The generator and predictor are trained jointly to find the minimal subgraph that maximizes prediction accuracy, enforcing an information bottleneck that discards irrelevant structural noise during the forward pass.

A fidelity assessment method that measures the change in a GNN's prediction after masking or altering the most important nodes or edges identified by an explainer. Common approaches include: removing top-k edges by importance score, adding random noise to critical node features, or comparing the original prediction distribution to the perturbed output. This technique directly validates GraphMask's claim that masked-away messages are truly irrelevant to the prediction.

A Layer-wise Relevance Propagation technique adapted for Graph Neural Networks that decomposes the output prediction backwards through the network to assign relevance scores to input nodes and edges. GNN-LRP follows conservative propagation rules that preserve the total relevance across layers. While GraphMask learns to suppress messages during the forward pass, GNN-LRP traces relevance backwards, offering a complementary perspective on information flow through the message-passing structure.