Research map / Learning theory and algorithms / Prediction and decision losses
Optimal Subgroup Discovery at Every Support Threshold
Lincen Yang, Qi Huang, Niki van Stein, Thomas Bäck, Matthijs van Leeuwen, Zhong Li
Read the paperCheck your idea against 137,217 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 an algorithm for subgroup discovery that addresses the deviance-support trade-off across all support thresholds rather than collapsing it into a single scalar objective.
Problem
Subgroup discovery seeks feature-space regions where a target deviates from its marginal distribution, but deviance and support trade off: small subgroups can be highly deviant yet statistically meaningless, while large ones cannot deviate much.
Approach
Existing methods collapse the deviance-support tension into a single scalar quality measure; this work instead targets optimal subgroups at every support threshold, so the full trade-off is recovered rather than one point on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| select atoms from a learned partition instead of folding support into a single score | subgroup discovery · 19 UCI datasets | average rank: 3.13 -> 1.25 | Improved |
| apply AtomSD atom-union refinement as post-processing to RSD, SD-KL, SD-mu, SyFlow subgroups (ablation) | competitor subgroups · 19 UCI datasets, 5 thresholds | held-out KL divergence | Improved |
| raise support threshold alpha from small to large | competitors (RSD, SD-KL, SD-mu, SyFlow) · 19 UCI datasets | valid subgroups returned | Hurt |
| raise support threshold alpha | AtomSD · 19 UCI datasets | valid subgroups returned | No clear effect |
| use support-weighted KL scoring with small subgroup samples | support-weighted KL · synthetic Gaussian mixture | ranking error | Hurt |
| evaluate KL on held-out test set instead of training set (ablation) | KL estimate · all subgroups, bootstrap | KL variance | No clear effect |
| use ratio-based support thresholds instead of absolute counts | AtomSD · 19 UCI datasets | held-out KL divergence | No clear effect |
| add a CDTree partition on small datasets (ForestFires) | AtomSD · ForestFires | held-out KL divergence: 0.000 vs SyFlow 0.579 | Hurt |
Closest papers by meaning
- Self-Organized Conformal Prediction: Reducing Regional Coverage Gaps with Unsupervised Group Discovery (NeurIPS 2026 · similarity 0.85)
- Subgroups Matter for Robust Bias Mitigation (ICML 2025 · similarity 0.85)
- Discover and Mitigate Multiple Biased Subgroups in Image Classifiers (CVPR 2024 · similarity 0.84)
- A Unified Framework for Variable Selection in Model-Based Clustering with Missing Not at Random (NeurIPS 2025 · similarity 0.84)
- Learning With Multi-Group Guarantees For Clusterable Subpopulations (ICML 2025 · similarity 0.84)
- Stochastic Grouping Conformal Prediction for Effective Subgroup Reliability (NeurIPS 2026 · similarity 0.84)
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.
