# Iterative Counterfactual Data Augmentation

arXiv: https://arxiv.org/abs/2502.18249 · HTML: https://mlsanity.com/paper/2502.18249/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 8 of its 18 references that are in the map.
- First posted on arXiv 25 Feb 2025.

## Builds on (most cited first, 8 of 8)

- [Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation](https://mlsanity.com/paper/1308.3432/) (2013)
- [Rationalizing Neural Predictions](https://mlsanity.com/paper/1606.04155/) (2016)
- [Learning to Explain: An Information-Theoretic Perspective on Model Interpretation](https://mlsanity.com/paper/1802.07814/) (2018)
- [Gender Bias in Neural Natural Language Processing](https://mlsanity.com/paper/1807.11714/) (2018)
- [Hierarchical Transformers for Long Document Classification](https://mlsanity.com/paper/1910.10781/) (2019)
- [Prompting Large Language Models for Counterfactual Generation: An Empirical Study](https://mlsanity.com/paper/2305.14791/) (2023)
- [D-Separation for Causal Self-Explanation](https://mlsanity.com/paper/2309.13391/) (2023)
- [Making a (Counterfactual) Difference One Rationale at a Time](https://mlsanity.com/paper/2201.05177/) (2022)

## Abstract (arXiv)

Counterfactual data augmentation (CDA) is a method for controlling information or biases in training datasets by generating a complementary dataset with typically opposing biases. Prior work often either relies on hand-crafted rules or algorithmic CDA methods which can leave unwanted information in the augmented dataset. In this work, we show iterative CDA (ICDA) with initial, high-noise interventions can converge to a state with significantly lower noise. Our ICDA procedure produces a dataset where one target signal in the training dataset maintains high mutual information with a corresponding label and the information of spurious signals are reduced. We show training on the augmented datasets produces rationales on documents that better align with human annotation. Our experiments include six human produced datasets and two large-language model generated datasets.