Research map / Label-efficient and robust learning / Recognition and long-tail
Iterative Counterfactual Data Augmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes iterative counterfactual data augmentation (ICDA), which repeatedly uses rationale selectors to shuffle sentence rationales across documents and retrain, converging to rationales that better match human annotations on beer, hotel, and LLM-generated review datasets.
Problem
Counterfactual data augmentation for rationale networks is limited by noisy interventions from an initial rationale selector that often picks spurious signals, so the augmented data still carries unwanted information and rationales misalign with human annotation.
Approach
Apply CDA iteratively: each round trains a new rationale selector on the augmented data and uses it to generate lower-noise counterfactuals for the next round, forming a fixed-point process that converges toward the maximum mutual information rationale; counterfactuals are made by shuffling sentence rationales between documents of opposite labels.
Key results
- On RateBeer smell, ICDA reaches 93.6 rationale precision vs 56.0 for MMI, 77.8 for CDA, and 50.8 for COMP.
- On LLM-generated Connectivity data, ICDA achieves 63.2 rationale precision vs 50.3 for MMI, 57.9 for CDA, and 56.9 for COMP.
- ICDA outperforms all baselines on the averaged results across all eight experiments, though on Restaurant CDA averaged below MMI due to one of three runs converging to a non-food aspect.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Continue counterfactual data augmentation iteratively, using the newly trained rationale selector at each iteration | CDA rationale models · RateBeer Appearance test | rationale precision: 62.9 -> 66.6 | Helped |
| Continue counterfactual data augmentation iteratively, using the newly trained rationale selector at each iteration | CDA rationale models · RateBeer Smell test | rationale precision: 77.8 -> 93.6 | Helped |
| Continue counterfactual data augmentation iteratively, using the newly trained rationale selector at each iteration | CDA rationale models · RateBeer Palatability test | rationale precision: 24.7 -> 37.6 | Helped |
| Continue counterfactual data augmentation iteratively, using the newly trained rationale selector at each iteration | CDA rationale models · TripAdvisor Location test | rationale precision: 84.7 -> 88.0 | Helped |
| Continue counterfactual data augmentation iteratively, using the newly trained rationale selector at each iteration | CDA rationale models · TripAdvisor Service test | rationale precision: 70.0 -> 74.0 | Helped |
| Continue counterfactual data augmentation iteratively, using the newly trained rationale selector at each iteration | CDA rationale models · TripAdvisor Cleanliness test | rationale precision: 53.9 -> 56.0 | Helped |
| Continue counterfactual data augmentation iteratively, using the newly trained rationale selector at each iteration | CDA rationale models · Connectivity test | rationale precision: 57.9 -> 63.2 | Helped |
| Continue counterfactual data augmentation iteratively, using the newly trained rationale selector at each iteration | CDA rationale models · Restaurant test | rationale precision: 51.5 -> 57.4 | Helped |
| Add one counterfactual data augmentation iteration to MMI | MMI rationale models · RateBeer Appearance, Smell, and Palatability; TripAdvisor Location, Service, and Cleanliness | rationale precision: 47.5 -> 62.9; 56.0 -> 77.8; 23.1 -> 24.7; 71.0 -> 84.7; 66.8 -> 70.0; 55.8 -> 53.9 | Mixed |
| Add one counterfactual data augmentation iteration to MMI | MMI rationale models · Connectivity and Restaurant test | rationale precision: 50.3 -> 57.9; 54.6 -> 51.5 | Mixed |
Surprising findings
- Deliberately degenerate counterfactual generation that injects the target-aspect rationale regardless of context can be information-theoretically beneficial, since increasing the helpful error rate beta raises CDA's error budget.
- Restaurant experiments converged to the Food aspect in only 2 of 3 runs, causing average CDA performance to drop below MMI.
Limitations
- ICDA costs multiple training runs per iteration, making it substantially more expensive than the baselines.
- Rationale training is stochastic and unstable (e.g., degeneration and MMI variance across seeds), which the binary theory does not capture.
- Restaurant dataset sometimes converges to a different aspect than the intended one, hurting averaged results.
Experimental setup
Datasets: RateBeer (correlated splits) TripAdvisor hotel reviews LLM-generated Connectivity headphone reviews LLM-generated Restaurant reviews
Benchmarks: RateBeer human rationale annotations (appearance, smell, palate) TripAdvisor human rationale annotations (location, service, cleanliness) Connectivity and Restaurant LLM-generated datasets
Metrics: rationale precision
Baselines: MMI COMP CDA
Builds on: rationale networks (Lei et al.) maximum mutual information rationalization counterfactual data augmentation (Kaushik et al.) complement control (COMP) CDA for rationales
Closest papers by meaning
- Making a (Counterfactual) Difference One Rationale at a Time (NeurIPS 2021 · similarity 0.89)
- Counterfactual Contrastive Learning with Normalizing Flows for Robust Treatment Effect Estimation (ICML 2025 · similarity 0.81)
- ACAMDA: Improving Data Efficiency in Reinforcement Learning through Guided Counterfactual Data Augmentation (AAAI 2024 · similarity 0.81)
- Exogenous Matching: Learning Good Proposals for Tractable Counterfactual Estimation (NeurIPS 2024 · similarity 0.81)
- Causal Action Influence Aware Counterfactual Data Augmentation (ICML 2024 · similarity 0.81)
- Conditional Counterfactual Mean Embeddings: Doubly Robust Estimation and Learning Rates (NeurIPS 2026 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.