Counterfactual Data Augmentation

Counterfactual data augmentation studies how to create training examples that change a meaningful factor—such as an image region, sentence meaning, or object position—while keeping other relevant information stable. Common approaches use generative models, language models, causal models, or minimal rule-based edits, then train with contrastive, supervised, or preference-based objectives. The methods are used to address scarce or imbalanced data, reduce reliance on spurious correlations, and improve robustness, fairness, and generalization in settings from medical imaging to robotics and language. A central challenge is ensuring that edits are realistic and change only the intended factor; poorly controlled examples or excessive augmentation can hurt performance, so validation and augmentation dose remain important.