Research map / Probabilistic and scientific modeling / Molecules and proteins
Generating Counterfactual Explanations Under Temporal Constraints
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 temporally constrained counterfactual generation for process traces by embedding LTLp automata (DFA) into a genetic algorithm's crossover and mutation operators so outputs satisfy background knowledge by design.
Problem
Counterfactual explanations for process traces (temporal data) often violate temporal background knowledge expressed in LTLp, yielding inconsistent, unrealistic explanations.
Approach
Infuse automata-theoretic (DFA) techniques for LTLp into a genetic algorithm: knowledge-aware crossover keeps constrained activities fixed, and two mutation strategies (FirstHeur, SecondHeur using SafeAct on the DFA) only apply mutations that preserve satisfaction, with proofs of correctness.
Key results
- Note: The above has a formatting error; corrected in the call below.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the first temporal knowledge-aware mutation strategy instead of a standard GA | Counterfactual generation with XGBoost · BPIC2017, 50% coverage | Distance: 0.6 -> 0.09 | Helped |
| Use the first temporal knowledge-aware mutation strategy instead of a standard GA | Counterfactual generation with XGBoost · BPIC2017, 50% coverage | Sparsity: 7.11 -> 1.06 | Helped |
| Use the first temporal knowledge-aware mutation strategy instead of a standard GA | Counterfactual generation with XGBoost · BPIC2012, 50% coverage | Runtime: 535 -> 451 | Helped |
| Use the first temporal knowledge-aware mutation strategy instead of a standard GA | Counterfactual generation with XGBoost · Claim Management, 50% coverage | Runtime: 3.39 -> 4.80 | Hurt |
| Use the DFA-guided second temporal knowledge-aware mutation strategy instead of a standard GA | Counterfactual generation with XGBoost · BPIC2012, 50% coverage | Distance: 0.55 -> 0.30 | Helped |
| Use the DFA-guided second temporal knowledge-aware mutation strategy instead of a standard GA | Counterfactual generation with XGBoost · BPIC2012, 50% coverage | Runtime: 535 -> 780 | Hurt |
| Use the DFA-guided second temporal knowledge-aware mutation strategy instead of a standard GA | Counterfactual generation with XGBoost · BPIC2017, 50% coverage | Sparsity: 7.11 -> 2.16 | Helped |
| Use temporal knowledge-aware crossover and mutation to preserve LTLp constraints | Proposed counterfactual generation methods · Process traces in the evaluation | Temporal-constraint compliance: 80% to 99% | Helped |
Surprising findings
- The standard GA (LTLf-GA) is often fastest in runtime, and SecondHeur is slower than the baseline at higher coverage.
- On the synthetic Claim Management dataset at low coverage, the standard GA shows better counterfactual quality than the proposed methods.
Limitations
- Standard GA's compliance is not guaranteed and is only measured empirically (80-99%).
- SecondHeur runtime grows with coverage and is higher than baselines in several settings.
- Only activity-level LTLp is handled; numerical data are left to future work.
Experimental setup
Datasets: Claim Management BPIC2012 BPIC2017
Benchmarks: Claim Management BPIC2012 BPIC2017
Metrics: distance sparsity implausibility diversity runtime compliance
Baselines: LTLf-GA (standard genetic algorithm) Mutate-And-Retry (MAR)
Builds on: genetic algorithm counterfactual generation LTLf LTLp
Closest papers by meaning
- DiCoFlex: Model-Agnostic Diverse Counterfactuals with Flexible Control (NeurIPS 2025 · similarity 0.87)
- Counterfactual Explanations of Time Varying Rankings (Student Abstract) (AAAI 2025 · similarity 0.87)
- Robust Stochastic Graph Generator for Counterfactual Explanations (AAAI 2024 · similarity 0.86)
- LeapFactual: Reliable Visual Counterfactual Explanation Using Conditional Flow Matching (NeurIPS 2025 · similarity 0.86)
- Counterfactual reasoning: an analysis of in-context emergence (NeurIPS 2025 · similarity 0.86)
- Generating Likely Counterfactuals Using Sum-Product Networks (ICLR 2025 · 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.