counterfactual data augmentation

**Counterfactual data augmentation** is the **fairness method that generates paired training examples by changing protected attributes while preserving task semantics** - CDA reduces spurious correlations learned from imbalanced data. **What Is Counterfactual data augmentation?** - **Definition**: Creation of counterfactual samples where identity terms are swapped and labels remain logically consistent. - **Goal**: Encourage models to treat protected attributes as irrelevant for neutral tasks. - **Common Transformations**: Pronoun swaps, name substitutions, and role-attribute replacements. - **Quality Requirement**: Counterfactuals must remain grammatically correct and semantically valid. **Why Counterfactual data augmentation Matters** - **Correlation Symmetry**: Breaks one-sided associations embedded in raw training corpora. - **Fairness Gains**: Often reduces demographic disparities in model predictions and generations. - **Data Efficiency**: Improves fairness without collecting entirely new datasets from scratch. - **Mitigation Flexibility**: Can target specific bias axes with controllable transformation rules. - **Benchmark Performance**: Frequently improves outcomes on stereotype bias evaluations. **How It Is Used in Practice** - **Transformation Rules**: Define safe attribute swaps with grammar-aware constraints. - **Label Preservation Checks**: Verify augmented pairs maintain correct task labels. - **Training Integration**: Mix original and counterfactual data with balanced sampling policy. Counterfactual data augmentation is **a practical and widely used fairness intervention** - well-constructed counterfactual pairs can materially reduce learned stereotype bias in language models.