counterfactual explanations

Counterfactual explanations show minimal input changes that would flip the model's decision. **Format**: "If X had been different, prediction would change from A to B." More actionable than feature importance. **Example**: Loan denial → "If income were $5K higher, loan would be approved." **Finding counterfactuals**: Optimization to find minimal edit that changes prediction, generative models to produce realistic alternatives, search over discrete changes (for text). **Desirable properties**: Minimal change (sparse, plausible), proximity to original, achievable/realistic, diverse set of counterfactuals. **For text**: Token substitutions, insertions, deletions that change classification. Challenge: maintaining fluency and semantic plausibility. **Advantages**: Actionable insights, intuitively understandable, recourse guidance. **Challenges**: Multiple valid counterfactuals exist, may suggest unrealistic changes, computationally expensive to find optimal. **Applications**: Lending/credit decisions, hiring, medical diagnosis, moderation appeal. **Tools**: DiCE, Alibi, custom search algorithms. **Regulatory relevance**: GDPR "right to explanation" - counterfactuals provide meaningful explanation of decisions. Powerful for high-stakes decisions.