From Abductive Explanations to Global Logical Rules for Node Classification in SGCs
Proposes a logic-based framework for explaining node classification in Simple Graph Convolution (SGC) networks using minimal abductive explanations as an intermediate representation Computes minimal sets of node-feature pairs sufficient to preserve predicted classes, addressing redundancy issues in existing subgraph-based approaches like LogicXGNN Trains decision trees from abductive explanations to extract compact global logical rules with high fidelity to the original SGC model Demonstrates ef
Analysis
TL;DR
- Proposes a logic-based framework for explaining node classification in Simple Graph Convolution (SGC) networks using minimal abductive explanations as an intermediate representation
- Computes minimal sets of node-feature pairs sufficient to preserve predicted classes, addressing redundancy issues in existing subgraph-based approaches like LogicXGNN
- Trains decision trees from abductive explanations to extract compact global logical rules with high fidelity to the original SGC model
- Demonstrates effectiveness on benchmark datasets, producing more generalizable rules by avoiding node-specific structural redundancy
Why It Matters
This work addresses a critical gap in GNN interpretability by moving beyond local subgraph explanations toward globally applicable logical rules, enabling practitioners to understand and trust SGC predictions at scale. The approach bridges abductive reasoning and rule extraction, offering a principled path to transparent graph-based machine learning systems that can be deployed in high-stakes domains requiring explainability.
Technical Details
- Abductive Explanation Computation: For each node, the framework identifies a minimal set of node-feature pairs that are sufficient to preserve the SGC model's predicted class, stripping away redundant structural information specific to individual nodes
- Decision Tree Training: The collected abductive explanations serve as training data for decision trees, which act as an intermediate symbolic representation bridging local explanations and global rules
- Global Rule Extraction: Logical rules are derived from the trained decision trees, producing compact, interpretable rules that generalize across the graph rather than being tied to individual node neighborhoods
- SGC-Specific Design: The method is tailored for Simple Graph Convolution networks, leveraging their simplified message-passing architecture to make abductive explanation computation tractable
- Benchmark Evaluation: Experiments on standard benchmark datasets validate that the extracted rules maintain high fidelity to the original SGC model while achieving greater compactness compared to prior subgraph-based approaches
Industry Insight
- The shift from subgraph-based to abductive explanation-based rule extraction represents a meaningful advancement in GNN interpretability, suggesting that future XGNN tools should prioritize feature-level minimality over structural neighborhood analysis
- As regulatory pressure for AI explainability grows, logic-based rule extraction from GNNs like SGC offers a compliance-friendly pathway for deploying graph models in regulated industries such as finance and healthcare
- The decision tree intermediary approach is broadly applicable beyond SGC and could be adapted to other GNN architectures, making it a versatile template for the emerging explainable GNN toolkit
Disclaimer: The above content is generated by AI and is for reference only.