Graph Explainability

A model-agnostic method that identifies the most influential subgraph and node features responsible for a specific prediction. It learns a compact subgraph and a feature mask by maximizing the mutual information between the original model's prediction and the prediction made using only the distilled information.

• Output: A minimal subgraph critical to the decision.
• Supply Chain Use: Pinpoints the exact supplier tier and specific attributes (e.g., lead time, credit score) that caused a delay prediction.

A parameterized, generative version of GNNExplainer that learns a global understanding of what constitutes a critical subgraph across the entire dataset. Instead of optimizing for a single instance, it trains a neural network to predict edge importance.

• Advantage: Provides explanations for new, unseen nodes without re-optimization.
• Supply Chain Use: Instantly identifies the critical path in a new Bill of Materials (BOM) graph that is likely to cause a production stoppage.

Adapts the classical Integrated Gradients method to graph data by attributing a prediction to the input features along a path from a neutral baseline to the actual input. It satisfies the completeness axiom, ensuring the sum of attributions equals the difference in output.

• Output: Node and edge feature attribution scores.
• Supply Chain Use: Quantifies the exact contribution of each warehouse's inventory level to a predicted stockout risk score.

A local interpretable model-agnostic explanation technique that approximates the GNN's decision boundary around a specific node using a simpler, interpretable model like a Hilbert-Schmidt Independence Criterion (HSIC) Lasso. It operates on the node's local neighborhood.

• Mechanism: Samples local perturbations and fits a linear model.
• Supply Chain Use: Explains why a specific logistics hub was flagged as a bottleneck by revealing the linear combination of neighboring transit times that triggered the alert.

A method that uses Monte Carlo Tree Search (MCTS) to efficiently explore the combinatorial space of subgraphs and identify the most explanatory substructure. It scores subgraphs using Shapley value approximations to ensure a fair distribution of contribution among nodes.

• Key Feature: Can identify non-contiguous, functional motifs.
• Supply Chain Use: Discovers that a specific combination of geographically dispersed but financially linked suppliers is the root cause of a systemic risk prediction.

Moves beyond node-level attribution to explain predictions in terms of high-level human-understandable concepts (e.g., 'bottleneck structure', 'hub-and-spoke'). It measures the alignment between latent graph representations and concept activation vectors.

• Mechanism: Tests if a prediction is sensitive to a specific conceptual pattern.
• Supply Chain Use: Confirms that a disruption prediction was triggered by the presence of a 'single-source dependency' pattern rather than a spurious correlation.

The foundational deep learning architecture that operates directly on graph-structured data. GNNs capture complex dependencies between entities through iterative message passing, where nodes aggregate feature information from their neighbors. Explainability techniques like GNNExplainer are specifically designed to open the black box of these models, revealing which subgraphs and node features drove a particular prediction.

A powerful GNN variant that employs self-attention mechanisms to assign different importance weights to neighboring nodes during feature aggregation. This inherent weighting makes GATs more naturally interpretable than standard GCNs, as the learned attention coefficients can be directly inspected to understand which neighbors the model considers most influential for a given node's prediction.

The core computational mechanism in most GNNs where nodes iteratively aggregate feature information from their local neighborhood and update their own hidden state. Explainability methods often work by identifying which specific messages (edges) were most critical during this process. By masking or perturbing these message pathways, techniques like GNNExplainer isolate the computational subgraph responsible for a model's output.

A low-dimensional, dense vector representation that encodes a node's structural position and feature information within the graph. Explainability techniques often operate in this latent space, identifying which dimensions of the embedding are most activated by specific input features. Understanding node embeddings is critical for interpreting why two nodes are classified similarly or dissimilarly by the model.

A theoretically maximally powerful GNN architecture designed to be as discriminative as the Weisfeiler-Lehman (WL) graph isomorphism test. Because GIN's aggregation function is injective, it captures fine-grained structural differences. Explaining a GIN model often involves identifying the specific subtree structures that differentiate one node's neighborhood from another, providing a rigorous structural basis for predictions.

A paradigm that jointly learns an optimal graph topology and node representations when the underlying graph is noisy, incomplete, or entirely absent. Explainability in this context is dual-purpose: it must clarify both why a specific latent graph structure was inferred and how that learned structure influences downstream predictions. This is crucial for supply chains where hidden dependencies may exist.