Logic-Based Framework Extracts Global Logical Rules from Abductive Explanations for Node Classification in SGCs
A novel framework grounded in logic seeks to enhance the interpretability of Graph Neural Networks (GNNs) specifically for node classification by generating global logical rules from minimal abductive explanations. This research, available as preprint arXiv:2608.17103, centers on Simple Graph Convolution (SGC) networks, a streamlined type of GNN. The method identifies a minimal collection of node-feature pairs for each node that is adequate to maintain the predicted class. These explanations act as an intermediary for training decision trees, facilitating the extraction of global logical rules. This approach overcomes the shortcomings of current logic-based explainers like LogicXGNN, which often rely on redundant structural information from explanatory subgraphs. The study includes experiments on benchmark datasets, although specific outcomes are not disclosed. This work contributes to the expanding domain of interpretable artificial intelligence, striving to clarify GNN predictions.
Key facts
- The framework is designed for node classification in Simple Graph Convolution (SGC) networks.
- Minimal abductive explanations are used as an intermediate representation for rule extraction.
- For each node, a minimal set of node-feature pairs sufficient to preserve the predicted class is computed.
- Decision trees are trained from these explanations to extract global logical rules.
- The method addresses limitations of LogicXGNN, which uses explanatory subgraphs with potentially redundant structural information.
- The paper is a preprint on arXiv with identifier 2608.17103.
- Experiments are conducted on benchmark datasets, though specific results are not provided in the available content.
Entities
Institutions
- arXiv