Research map / Label-efficient and robust learning / Recognition and long-tail
Counterfactual Explanations for Misclassified Images: How Human and Machine Explanations Differ (Abstract Reprint)
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.
Collects human-generated counterfactual edits for CNN misclassifications on MNIST and QuickDraw and compares them with four counterfactual XAI methods on proximity, representativeness, and prototypicality metrics.
Problem
Whether counterfactual XAI methods for image classifiers generate explanations that people find plausible is untested, since few counterfactual methods have been user-tested.
Approach
Instead of showing machine counterfactuals to users, ask people to edit misclassified images to produce counterfactual explanations, then compare these human ground-truth edits with machine outputs from Min-Edit, CEM-PN, VLK and Revise using distance, out-of-distribution and prototype-similarity metrics.
Key results
- Human counterfactual edits differ from machine methods: Min-Edit, CEM-PN and VLK have significantly lower L1 and L2 distances than human edits on MNIST (all p<0.001), indicating humans do not minimally edit.
- Human counterfactuals score highest on representativeness measures, e.g., MC-Mean 0.94 vs 0.33-0.66 for machine methods and R%-Sub 50.05 on MNIST.
- Human counterfactuals have higher Grad-Cos similarity to counterfactual-class MMD-Critic prototypes than all four XAI methods on both datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use human-generated counterfactuals instead of machine-generated explanations | Counterfactual explanations · MNIST and QuickDraw | proximity to counterfactual-class prototypes | Helped |
| Use human-generated counterfactuals instead of machine-generated explanations | Counterfactual explanations · MNIST and QuickDraw | R%-Sub: 50.05; 55.98 | Helped |
| Use human-generated counterfactuals instead of machine-generated explanations | Counterfactual explanations · MNIST | MC-Mean: 0.94 | Helped |
| Use human-generated counterfactuals instead of machine-generated explanations | Counterfactual explanations · QuickDraw | 10-LOF: 0.05 vs 0.12 | Hurt |
| Instruct participants to make the smallest possible edits (ablation) | Human MNIST counterfactuals · MNIST | L1 and L2 distance: p < 0.001 | Helped |
| Instruct participants to make the smallest possible edits (ablation) | Human MNIST counterfactuals · MNIST | L1 and L2 distance versus machine-generated explanations: all p < 0.001 | No clear effect |
Surprising findings
- Instructing participants to make the smallest possible changes reduced their edit distances but they still made larger edits than the machine methods.
- Humans tended to add pixels more often than remove them, adding semantic features such as rotor blades to helicopters.
Limitations
- Small sample for QuickDraw (N=5 pilot) and 50 MNIST / 30 QuickDraw items, limiting statistical power.
- Participants were Computer Science students, not a general population.
- Revise failed to generate counterfactuals for many QuickDraw instances, so its results reflect a subset.
Experimental setup
Datasets: MNIST QuickDraw Doodles
Benchmarks: MNIST QuickDraw Doodles
Metrics: L1 distance L2 distance MC-Dropout mean MC-Dropout std IM1 10-LOF R%-Substitutability Grad-Cos similarity
Baselines: Min-Edit CEM-PN VLK Revise
Builds on: Wachter et al. counterfactual method CEM-PN VLK Revise MMD-Critic
Closest papers by meaning
- V-CECE: Visual Counterfactual Explanations via Conceptual Edits (NeurIPS 2025 · similarity 0.91)
- Towards Unifying Evaluation of Counterfactual Explanations: Leveraging Large Language Models for Human-Centric Assessments (AAAI 2025 · similarity 0.90)
- LeapFactual: Reliable Visual Counterfactual Explanation Using Conditional Flow Matching (NeurIPS 2025 · similarity 0.89)
- I-CEE: Tailoring Explanations of Image Classification Models to User Expertise (AAAI 2024 · similarity 0.88)
- Rethinking Visual Counterfactual Explanations Through Region Constraint (ICLR 2025 · similarity 0.88)
- Promoting Counterfactual Robustness through Diversity (AAAI 2024 · similarity 0.88)
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.