Machine-learning models are prone to capturing the spurious correlations
between non-causal attributes and classes, with counterfactual data
augmentation being a promising direction for breaking these spurious
associations. However, explicitly generating counterfactual data is
challeng