Research map / Learning theory and algorithms / Prediction and decision losses
Counterfactual Explanations Can Be Manipulated
Dylan Z Slack, Sophie Hilgard, Himabindu Lakkaraju, Sameer Singh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add learned perturbation delta before counterfactual search | Wachter et al. on manipulated model · Communities and Crime | cost reduction: 20.1x | Improved |
| add learned perturbation delta before counterfactual search | DiCE on manipulated model · Communities and Crime | recourse cost: 42.63 -> 9.57 | Improved |
| add learned perturbation delta before counterfactual search | Sparse Wachter on manipulated model · Communities and Crime | recourse cost: 52.05 -> 22.59 | Improved |
| add learned perturbation delta before counterfactual search | Prototypes on manipulated model · German Credit | recourse cost: 13.98 -> 4.69 | Improved |
| add delta to non-protected group on unmodified model | DiCE on unmodified model · Communities and Crime | recourse cost: 11.90 -> 11.50 | No clear effect |
| train manipulated model instead of unmodified model | Wachter et al. · Communities and Crime | recourse cost disparity: 3.59 -> 0.37 | Improved |
| train manipulated model instead of unmodified model | DiCE · Communities and Crime | recourse cost disparity: 24.41 -> 6.99 | Improved |
| train manipulated model instead of unmodified model | Wachter et al. · Communities and Crime | test accuracy: 81.2 -> 80.9 | No clear effect |
| train manipulated model instead of unmodified model | Sparse Wachter · Communities and Crime | test accuracy: 81.2 -> 77.9 | Hurt |
| move from Communities and Crime (99 features) to German Credit (7 features) | manipulated models · both datasets | cost reduction: 20.1x -> 1.8x | Hurt |
| round delta for categorical features instead of numerical-only delta (ablation) | Prototypes manipulated model · German Credit | cost reduction: 2.3x -> 0.92x | Hurt |
| initialize counterfactual search at x plus N(0,1) noise (ablation) | Wachter and DiCE counterfactual search · Communities and Crime | cost reduction | Hurt |
Closest papers by meaning
- Promoting Counterfactual Robustness through Diversity (AAAI 2024 · similarity 0.88)
- Collective Counterfactual Explanations: Balancing Individual Goals and Collective Dynamics (NeurIPS 2025 · similarity 0.87)
- A New Approach to Backtracking Counterfactual Explanations: A Unified Causal Framework for Efficient Model Interpretability (ICML 2025 · similarity 0.87)
- Generating Likely Counterfactuals Using Sum-Product Networks (ICLR 2025 · similarity 0.86)
- Counterfactual Identification Under Monotonicity Constraints (AAAI 2025 · similarity 0.86)
- Natural Counterfactuals With Necessary Backtracking (NeurIPS 2024 · similarity 0.86)
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.
