Lay Summary

Explainable AI aims to help humans understand why a machine learning model makes a specific prediction. One effective method is counterfactual explanations, which show the smallest change needed in input data to alter an outcome. However, generating these explanations for complex, graph-structured data—like social networks or chemical molecules—is challenging because structural rules must be preserved. This paper proposes a framework for creating “semantic graph counterfactuals.” By ensuring generated modifications remain realistic and logical within the graph’s context, the approach makes complex AI predictions more transparent and interpretable. (88 words)

Full Abstract

Counterfactual explanations have emerged as a prominent tool for interpreting complex machine learning decisions by highlighting minimal recourse modifications to inputs. However, extending counterfactual generation to graph-structured data presents distinct challenges due to the interconnected nature of node features and topological constraints. Existing graph counterfactual methods often produce non-sensical or structurally invalid modifications. We propose a novel framework for generating Semantic Graph Counterfactuals that respects semantic relationships and domain-specific structural invariants. By constraining node and edge perturbations using domain semantics, our framework outputs realistic, actionable, and interpretable counterfactual graphs, advancing explainability in graph neural network applications.

Metadata

• Publication Date: 2024
• Author: Angeliki Dimitriou, Maria Lymperaiou, Georgios Filandrianos, Konstantinos Thomas, Giorgos Stamou
• Conference / Journal: Machine Learning and Knowledge Extraction / Scientific Proceedings