Graph Rationalization

The architecture consists of two tightly coupled modules trained end-to-end:

• Generator Module: Learns a stochastic masking policy over edges or nodes to extract a compact subgraph—the rationale—from the full input graph.
• Predictor Module: Takes only the extracted rationale as input and performs the downstream task (classification, regression).

The generator is optimized to produce rationales that are both predictive (enable accurate task performance) and concise (minimize spurious correlations). This cooperative training ensures the rationale captures the true causal structure rather than superficial statistical shortcuts.

Graph Rationalization formalizes explanation as an information bottleneck optimization problem:

• Objective: Maximize mutual information between the extracted rationale and the label while minimizing mutual information between the rationale and the original graph.
• Effect: The generator is forced to discard non-causal, spurious correlations and retain only the minimal structural patterns that genuinely determine the prediction.
• Implementation: Often achieved through variational bounds, Gumbel-Softmax reparameterization for discrete edge selection, or reinforcement learning with a sparsity reward.

This principle directly connects to the Graph Information Bottleneck and GSAT (Graph Stochastic Attention) methods, which formalize the trade-off between compression and prediction.

Unlike post-hoc explainers (GNNExplainer, SubgraphX) that approximate a trained model's behavior after the fact, Graph Rationalization produces inherently faithful explanations:

• The predictor never sees the full graph during inference—it operates exclusively on the rationale.
• The explanation is the decision boundary; there is no gap between what the model uses and what is shown.
• This eliminates the faithfulness gap where post-hoc methods may highlight features the model did not actually rely on.

This property makes rationalization particularly suitable for high-stakes domains like drug discovery and fraud detection, where explanation fidelity is non-negotiable.

Selecting a discrete subgraph is a non-differentiable operation. Graph Rationalization employs several techniques to enable gradient-based optimization:

• Gumbel-Softmax Reparameterization: Provides a continuous relaxation of discrete edge sampling, allowing gradients to flow through the generator during training.
• Hard Concrete Distributions: Learn sparse masks with exact zeros, producing truly discrete rationales at inference time.
• REINFORCE with Baselines: Uses policy gradient methods from reinforcement learning to train the generator when discrete sampling is required, with a moving average baseline to reduce variance.
• Straight-Through Estimators: Approximate gradients through discrete sampling operations during backpropagation.

A critical advantage of Graph Rationalization is its ability to mitigate spurious correlations through causal intervention:

• The generator learns to do(X = rationale) rather than passively observe correlations, mimicking causal intervention in a Structural Causal Model.
• By training the predictor exclusively on extracted rationales, the model becomes robust to distribution shifts—changes in spurious patterns that would fool a standard GNN.
• DIR (Discovering Invariant Rationales) extends this by enforcing invariance across multiple training environments, ensuring the rationale captures stable causal mechanisms.

This connects Graph Rationalization to the broader field of invariant risk minimization and out-of-distribution generalization.

Graph Rationalization models are evaluated on three key axes:

• Prediction Fidelity: Accuracy of the predictor when using only the extracted rationale versus the full graph. Higher fidelity indicates the rationale captures all task-relevant information.
• Sparsity: The size of the rationale relative to the original graph, often measured as the percentage of retained edges. Optimal rationales are minimal but sufficient.
• Stability: Consistency of the extracted rationale under small input perturbations. A stable generator produces similar rationales for semantically similar graphs.

Additional metrics include contrastive accuracy (does the rationale alone enable correct classification?) and causal effect estimation (does intervening on the rationale change the prediction as expected?).

In pharmaceutical research, a graph rationalization model can identify the specific pharmacophore—the minimal set of atoms and bonds responsible for a molecule's therapeutic effect. By training a generator to extract the causal subgraph from a molecular graph, chemists can validate whether the model's prediction of toxicity or efficacy is based on known biochemical principles or spurious structural correlations. This directly accelerates lead optimization by focusing synthesis efforts on the functional core of a molecule.

Graph rationalization isolates the minimal fraudulent motif within a massive transaction graph. Instead of flagging an entire account, the generator extracts the specific sequence of transactions and counterparties that constitute the fraud signal. This allows compliance officers to audit the model's logic, ensuring it is focusing on structural patterns like layering or rapid fund dispersion rather than biased demographic proxies, satisfying regulatory requirements for adverse action reasoning.

When classifying neurological disorders from functional connectivity graphs, rationalization extracts the disease-specific sub-network. The generator identifies the critical neural pathways and brain regions whose connectivity patterns drive the diagnosis, discarding irrelevant background activity. This provides neurologists with a verifiable, compact biomarker—a specific circuit rather than a whole-brain map—enabling trust in the AI's diagnostic suggestion and alignment with clinical literature.

In a user-item interaction graph, rationalization explains a recommendation by extracting the causal subgraph of past interactions. For a 'buy' prediction, the generator might isolate a specific path: User -> Co-purchased Item -> Category -> Target Item. This moves beyond collaborative filtering's 'users like you also bought' to a verifiable, multi-hop reasoning chain, allowing the platform to surface the exact historical behavior that triggered the recommendation.

Graph rationalization models process provenance graphs from system logs to pinpoint the precise attack sequence leading to a compromise. The generator extracts the minimal causal chain of process creations, file writes, and network connections that constitute the attack, filtering out benign background noise. This provides security analysts with a concise, high-fidelity incident report, dramatically reducing triage time by highlighting the exact root cause and lateral movement path.

For predicting material properties like band gap or formation energy from crystal graphs, rationalization identifies the decisive local atomic environment. The generator extracts the minimal cluster of atoms and bonds whose arrangement dictates the macroscopic property. This helps materials scientists validate that the model has learned physically meaningful coordination chemistry rather than dataset artifacts, guiding the targeted synthesis of novel materials with desired properties.

The foundational information-theoretic principle underlying graph rationalization. It formalizes the objective of compressing an input graph into a minimal subgraph (the rationale) that preserves maximal mutual information with the label while discarding irrelevant structural noise.

• Balances compression (minimize rationale size) against sufficiency (maximize predictive power)
• The generator-predictor framework directly optimizes the IB Lagrangian
• Prevents trivial solutions where the generator copies the entire input graph

A quantitative evaluation score measuring how accurately an extracted rationale reflects the true reasoning process of the GNN. It is typically assessed by the drop in prediction performance when the rationale is removed from the input.

• High faithfulness: removing the rationale causes a sharp accuracy decline
• Low faithfulness: the model relies on spurious correlations outside the rationale
• Complementary to fidelity, which measures how well the rationale alone reproduces the original prediction

An evaluation metric that measures how well an explanation mimics the original model's behavior. Calculated as the accuracy of the original GNN when evaluated solely on the extracted explanatory subgraph.

• Fidelity+: accuracy of the rationale alone vs. the full graph
• Fidelity-: accuracy when the rationale is removed, revealing reliance on non-rationale features
• High fidelity indicates the rationale captures the model's decision boundary

The process of inferring cause-and-effect relationships between variables in graph-structured data, moving beyond correlation to identify the true generative mechanisms. This provides robust explanations that remain valid under distribution shift.

• Uses structural causal models (SCMs) to represent interventions and counterfactuals
• Enables answering 'what if' questions about node or edge removal
• Critical for building rationales that generalize beyond training distributions

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. It directly tests whether the highlighted structures are causally relevant.

• Positive perturbation: removing rationale nodes should drop confidence in the predicted class
• Negative perturbation: removing non-rationale nodes should have minimal impact
• Provides a ground-truth-free evaluation when human annotations are unavailable

Learning node embeddings where each dimension corresponds to an independent, interpretable generative factor of the graph data. This enables fine-grained explanation of latent structures by isolating the specific factor responsible for a prediction.

• Each latent dimension captures a distinct semantic property (e.g., functional group in a molecule)
• Facilitates controlled generation of counterfactual graph structures
• Improves the generator module's ability to separate causal from non-causal components