On Formal Counterfactual Explanations

Impact 6 arXiv cs.AI

Summary

  • The paper bridges closest-world counterfactual explanations with formal XAI concepts like abductive (AXps) and contrastive explanations (CXps).
  • It characterizes closest-world counterfactuals as a disjunctive normal form over minimally distant counterexample instances.
  • The authors introduce conjunctive counterfactual explanations (CFXps) as feature literal conjunctions ensuring prediction change.
  • Structural results connect CXps, AXps, and CFXps, enabling modular CFXp computation using existing explanation machinery.
  • Benchmarks on tree ensembles and neural networks show efficient CFXp extraction requiring comparable feature changes to point-based methods.

Why it matters

By bridging closest-world semantics and formal XAI via conjunctive counterfactual explanations, this research provides a modular and efficient approach for reliable model explainability.

Who should care