Research map / Label-efficient and robust learning / Recognition and long-tail
Regional Explanations: Bridging Local and Global Variable Importance
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.
Shows that local Shapley Values and LIME assign importance to irrelevant features even in ideal piecewise-linear settings, and proposes R-LOCO, which clusters observations in feature-importance space and assigns each region's average LOCO importance as local attribution.
Problem
Local attribution methods (Local Shapley Values and LIME) can assign importance to features that neither affect the model output locally nor are statistically dependent on functionally relevant features, even under independence and exact computation.
Approach
Partition the input space into regions with homogeneous feature-importance profiles (LOCO-based conformity-score differences clustered via affinity propagation or other clustering), then assign each instance the average global LOCO importance of its region, bridging local and global explanations.
Key results
- On the 1st-order synthetic model, R-LOCO achieves 0.97 TP and 0.03 FP, versus 0.41 TP / 0.59 FP for Local SV and 0.53 TP / 0.47 FP for LIME.
- Non-important features receive mean attribution 0.02 for R-LOCO versus 0.12 for Local SV and LIME on the 1st-order model.
- On real datasets (Diabetes, California), R-LOCO outperforms baselines in zero-masking faithfulness experiments with XGBoost models, and is reported to be faster at inference.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use R-LOCO instead of Local Shapley Values | local feature attribution · 1st-order synthetic | TP: 0.41 -> 0.97 | Helped |
| use R-LOCO instead of LIME | local feature attribution · 1st-order synthetic | TP: 0.53 -> 0.97 | Helped |
| use R-LOCO instead of LOCO | local feature attribution · 1st-order synthetic | TP: 0.50 -> 0.97 | Helped |
| use R-LOCO instead of Local Shapley Values | local feature attribution · 2nd-order synthetic | TP: 0.30 -> 0.86 | Helped |
| use R-LOCO instead of LIME | local feature attribution · 2nd-order synthetic | TP: 0.43 -> 0.86 | Helped |
| use R-LOCO instead of LOCO | local feature attribution · 2nd-order synthetic | TP: 0.50 -> 0.86 | Helped |
| use R-LOCO instead of Local Shapley Values | local feature attribution · 2nd-order-interaction synthetic | TP: 0.44 -> 0.67 | Helped |
| use R-LOCO instead of LIME | local feature attribution · 2nd-order-interaction synthetic | TP: 0.44 -> 0.67 | Helped |
| use R-LOCO instead of LOCO | local feature attribution · 2nd-order-interaction synthetic | TP: 0.50 -> 0.67 | Helped |
| use optimal clusters based on model structure rather than default clustering (ablation) | R-LOCO · 1st-order synthetic | TP: 0.97 -> 1.00 | Helped |
| cluster in feature-importance space rather than input space (ablation) | R-LOCO · 1st-order synthetic | TP: 0.89 -> 0.97 | Helped |
| cluster in feature-importance space rather than input space (ablation) | R-LOCO · 2nd-order synthetic | TP: 0.74 -> 0.86 | Mixed |
Surprising findings
- Local Shapley Values and LIME perform almost identically to the global LOCO method on the piecewise-linear benchmark, in terms of identifying locally important features.
- LIME coefficients change sign for about 20% of features when bandwidth is halved on the German credit dataset.
Limitations
- Assumes each region of the model can be approximated by an additive structure.
- Performance depends heavily on clustering quality; contaminated clusters bias attributions.
- Main theoretical and controlled experiments assume independent features.
Experimental setup
Datasets: synthetic piecewise-linear data Diabetes California German credit
Benchmarks: synthetic 1st-order model synthetic 2nd-order model synthetic 2nd-order-interaction model Diabetes California
Metrics: true positives false positives attribution mass on non-important features performance drop under zero-masking top-k and bottom-k featu
Baselines: Local SHAP LIME LOCO FDT CSHAP
Builds on: LIME Local Shapley Values LOCO Shapley Effects
Closest papers by meaning
- Less is More: Fewer Interpretable Region via Submodular Subset Selection (ICLR 2024 · similarity 0.87)
- Data-faithful Feature Attribution: Mitigating Unobservable Confounders via Instrumental Variables (NeurIPS 2024 · similarity 0.86)
- Feature Attribution with Necessity and Sufficiency via Dual-stage Perturbation Test for Causal Explanation (ICML 2024 · similarity 0.86)
- Provably Better Explanations with Optimized Aggregation of Feature Attributions (ICML 2024 · similarity 0.85)
- Prospector Heads: Generalized Feature Attribution for Large Models & Data (ICML 2024 · similarity 0.85)
- Local vs. Global Interpretability: A Computational Complexity Perspective (ICML 2024 · similarity 0.85)
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.