Research map / Label-efficient and robust learning / Recognition and long-tail
CohEx: A Generalized Framework for Cohort Explanation
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 CohEx, a framework that converts data-driven local feature-importance explainers into cohort-level explanations by iteratively alternating supervised clustering (SRIDHCR) with recomputing cohort-restricted importances, plus new metrics for cohort explanation desiderata.
Problem
Cohort (regional) explanations, the middle ground between global and local explanations, are under-explored; existing cohort methods lack evaluation metrics and suffer from leakage of out-of-cohort information into local importances.
Approach
Iteratively recompute local importances using only samples within each cohort and re-run supervised clustering on feature values plus importance scores until the clustering loss stops decreasing, so explanations respect cohort locality; also defines metrics for generalizability, conciseness, disjointness, locality, and stability.
Key results
- On the synthetic patient task with LIME, CohEx achieves the lowest cohort locality loss (0.0052) and importance stability loss (0.0027) among compared methods, while REPID+GALE and CohEx have the lowest clustering penalties (0.0005 and 0.0006).
- On bike sharing (SHAP, XGBoost R^2 0.9453), CohEx reduces clustering penalty to 0.3196 vs 0.6641 for VINE and 1.1633 for REPID, and locality loss to 0.5034 vs 2.5280 for REPID.
- CohEx has lower cohort stability than REPID (e.g., 40.76% vs 47.37% on the patient task; 43.68% vs 100% on bike sharing), which the authors attribute to the trial-and-error nature of the method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CohEx instead of the strongest listed baseline | Cohort explanation with LIME · Synthetic Patient Classification | Cohort Locality: 0.0157 \pm 0.0050 -> 0.0052 \pm 0.0088 | Helped |
| Use CohEx instead of the strongest listed baseline | Cohort explanation with LIME · Synthetic Patient Classification | Importance Stability: 0.0131 \pm 0.0039 -> 0.0027 \pm 0.0033 | Helped |
| Use CohEx instead of Hierarchical | Cohort explanation with SHAP · Bike Sharing | Clustering Penalty: 0.4796 \pm 0.1173 -> 0.3196 \pm 0.1018 | Helped |
| Use CohEx instead of REPID | Cohort explanation with SHAP · Bike Sharing | Cohort Locality: 2.5280 \pm 0.0091 -> 0.5034 \pm 0.2214 | Helped |
| Use CohEx instead of REPID | Cohort explanation with SHAP · Bike Sharing | Importance Stability: 2.1123 \pm 0.0046 -> 0.1022 \pm 0.0886 | Helped |
| Use CohEx instead of REPID+GALE | Cohort explanation with DeepLIFT · MNIST | Clustering Penalty: 1.3290 \pm 0.0000 -> 1.3112 \pm 0.3037 | Helped |
| Use CohEx instead of Hierarchical | Cohort explanation with DeepLIFT · MNIST | Cohort Locality: 0.0127 \pm 0.0047 -> 0.0010 \pm 0.0019 | Helped |
| Use CohEx instead of Hierarchical | Cohort explanation with DeepLIFT · MNIST | Importance Stability: 0.0956 \pm 0.0475 -> 0.0210 \pm 0.0419 | Helped |
| Add iterative local-importance recomputation and reclustering (ablation) | Hierarchical cohort explanation · Synthetic Patient Classification | Cohort Locality: 0.0157 \pm 0.0050 -> 0.0052 \pm 0.0088 | Helped |
Surprising findings
- With vanilla SHAP as the base method on the synthetic patient task, CohEx performs worse than VINE and REPID on clustering penalty and cohort locality, which the authors attribute to SHAP values being less similar across samples.
- GALE re-weighting harms cohort locality, since it smooths importance over the whole dataset.
Limitations
- Cohort definitions are not stable across runs (lower ARI than REPID).
- High time complexity, especially for vision tasks with sampling-based explainers.
- Framework is limited to local data-driven explainers; non-data-driven methods fall back to a hierarchical scheme without locality guarantees.
Experimental setup
Datasets: Synthetic patient classification (generated) Bike sharing (Capital bikeshare, 2011-2012) MNIST
Benchmarks: Synthetic patient classification Bike sharing MNIST subset of 7 and 9 digits
Metrics: clustering penalty (generalizability and conciseness) cohort locality loss cohort stability (ARI) importance stability disjoint cohort property
Baselines: VINE VINE+GALE REPID REPID+GALE Hierarchical cohort explanation
Builds on: VINE REPID GALE LIME SHAP
Closest papers by meaning
- Model Agnostic Multilevel Explanations (NeurIPS 2020 · similarity 0.84)
- Local vs. Global Interpretability: A Computational Complexity Perspective (ICML 2024 · similarity 0.83)
- iLLuMinaTE: An LLM-XAI Framework Leveraging Social Science Explanation Theories Towards Actionable Student Performance Feedback (AAAI 2025 · similarity 0.83)
- Evaluate with the Inverse: Efficient Approximation of Latent Explanation Quality Distribution (AAAI 2025 · similarity 0.83)
- Position: Rethinking Explainable Machine Learning as Applied Statistics (ICML 2025 · similarity 0.83)
- Navigating the Maze of Explainable AI: A Systematic Approach to Evaluating Methods and Metrics (NeurIPS 2024 · similarity 0.83)
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.