Research map / Learning theory and algorithms / Prediction and decision losses
Learning Performance Maximizing Ensembles with Explainability Guarantees
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 Ensembles with Explainability Guarantees (EEG), which learns a global allocator ranking observations by glass-box allocation desirability so that a glass box model handles a chosen proportion q of inputs while keeping ensemble performance high.
Problem
Partial explainability: choosing which observations a glass box model should predict (at a given explainability level) versus a black box model, so that ensemble performance is maximized while explainability is maximized.
Approach
Independently learn glass box and black box models on the full data, then learn an allocator that ranks observations by a sufficiency-based allocation desirability percentile (combining sufficiency categories and relative loss); the ranking is constant in q, so one learned allocator yields optimal allocations for any explainability level.
Key results
- Learned allocation achieves cross-dataset PPCR of 37% over random allocation, with near-oracle performance on some datasets (89% on IsoletR).
- Allocated ensembles match or exceed the best component model across an average of 74% of the explainability range (PQEOM).
- Max accuracy averages 84% achieved at an average explainability level of 64% in the single-allocation setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| learn observation allocations between glass-box and black-box models | EEG · 31 tabular datasets | PPCR: 37 ± 21 | Helped |
| learn observation allocations between glass-box and black-box models | EEG · 31 tabular datasets | PQEOM: 74 ± 19 | Helped |
| learn observation allocations between glass-box and black-box models | EEG · 31 tabular datasets | 95TQM: 94 ± 9 | Helped |
| learn allocations for complementary component models | EEG · PolR | PQOM: 93 ± 1 | Helped |
| use a feature-dependent allocator alongside the feature-independent allocator | EEG · 31 tabular datasets | PCFA: 35 ± 34 | No clear effect |
| use only original features x for allocator training (ablation) | EEG allocator · Allocator feature-set study | datasets not significantly worse than best feature set (α=0.05): 18.75% | Hurt |
| use all candidate features x, g, b, dce and dmse (ablation) | EEG allocator · Allocator feature-set study | datasets not significantly worse than best feature set (α=0.05): 56.25% | Helped |
| vary the allocator feature set across candidate subsets (ablation) | EEG allocator · Allocator feature-set study | AUC | Mixed |
| select component models by ensemble validation performance instead of individual validation performance (ablation) | EEG · tabular datasets | AUC: 0.01 ± 0.03 | No clear effect |
| select component models by ensemble validation performance (ablation) | EEG · some datasets, including IsoletR and BrazilianHousesR | AUC: 15% | Helped |
Surprising findings
- Feature-dependent allocator adds no reliable benefit over the feature-independent distance-based allocator on average (PCFA 35% ± 34%).
- Selecting component models by ensemble performance rather than individual performance changed allocator performance by only 0.01 ± 0.03 on average.
Limitations
- Experiments limited to tabular data with a restricted set of glass box and black box model types.
- Feature-dependent allocator gives limited average gain over feature-independent allocation.
- Full model-combination and feature-set searches are costly, so simplified selection may be needed.
Experimental setup
Datasets: Grinsztajn et al. tabular benchmark (31 datasets)
Benchmarks: Grinsztajn tabular benchmark (31 datasets)
Metrics: AUC PPCR PQEOM PQOM PCFA 95TQM Max Acc Argmax q sufficiency accuracy
Baselines: Oracle allocator Random allocator Glass box models (logistic/linear regression, classification/regression tree) Black box models (gradient boosting trees, neural network)
Builds on: LSP OTSAM HyRS HybridCORELS Wide ResNet
Closest papers by meaning
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.84)
- Tree Ensemble Explainability through the Hoeffding Functional Decomposition and TreeHFD Algorithm (NeurIPS 2025 · similarity 0.83)
- What makes an Ensemble (Un) Interpretable? (ICML 2025 · similarity 0.83)
- Conformal Prediction for Ensembles: Improving Efficiency via Score-Based Aggregation (NeurIPS 2025 · similarity 0.83)
- Efficient and Accurate Explanation Estimation with Distribution Compression (ICLR 2025 · similarity 0.83)
- MEPSI: An MDL-Based Ensemble Pruning Approach with Structural Information (AAAI 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.