Research map / Label-efficient and robust learning / Recognition and long-tail

Learning under Localized Minority Imbalance

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Group-robust classification · 0 citations (OpenAlex)

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 localized minority imbalance (LMI), where minority observability depends on class and features, and a tree-based stratified ensembling method (SE) plus an unbiased-test evaluation protocol to handle it.

Problem

Class-imbalance methods assume minority instances are uniformly undersampled; under localized minority imbalance (LMI), observation bias depends on feature regions, so existing methods fit observation-induced bias and generalize poorly, and hold-out evaluation from biased data misleads.

Approach

A tree that recursively partitions feature space to minimize within-node LMI distortion, measured by parent-child changes in observed logit minority prevalence over shared bins; each stratum pairs its majority instances with all observed minority instances, and the stratum classifiers are ensembled with performance weights.

Key results

• SE attains the highest average MCC and F1 on undamaged test data across five base classifiers; e.g., with logistic regression MCC 0.57 vs 0.54 for the best baseline (ICF) and 0.12 for LA.
• With random forest, SE reaches MCC 0.59 versus 0.47 for LNG and 0.46 for RE.
• Conventional hold-out evaluation from LMI-biased data overstates performance of methods that overfit the observed minority subset, while gold-standard evaluation reverses rankings.

Surprising findings

• DMG (no imbalance correction) outperforms the gold-standard-data baseline in low-LMI regions, reflecting overfitting to observation bias.
• AUC metrics are much less informative under LMI, with all methods performing similarly.

Limitations

• LMI is not identifiable from observed data alone, so the SE objective is only an observed-data surrogate and cannot certify constant LMI distortion.
• Gold-standard evaluation requires unbiased test sets that are often unavailable; LMI experiments are simulated on balanced benchmarks.
• SE does not improve over existing methods under pure class-count imbalance.

Experimental setup

Datasets: page-blocks0 abalone19 yeast3 yeast4 yeast5 yeast6 segment0 shuttle-c0-vs-c4

Benchmarks: KEEL benchmark datasets with simulated LMI and class-count imbalance

Metrics: MCC F1 balanced accuracy PR AUC ROC AUC precision recall

Baselines: ICF FL MET LNG CLS LA RE EE BB CC TMK SMOTE

Builds on: SMOTE EasyEnsemble Balanced Bagging focal loss logit adjustment

Closest papers by meaning

• Balanced Data, Imbalanced Spectra: Unveiling Class Disparities with Spectral Imbalance (ICML 2024 · similarity 0.90)
• Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data (ICML 2025 · similarity 0.87)
• Concentration and excess risk bounds for imbalanced classification with synthetic oversampling (NeurIPS 2025 · similarity 0.87)
• CLIMB: Class-imbalanced Learning Benchmark on Tabular Data (NeurIPS 2025 · similarity 0.87)
• Restoring balance: principled under/oversampling of data for optimal classification (ICML 2024 · similarity 0.86)
• Synthetic Tabular Data Generation for Imbalanced Classification: The Surprising Effectiveness of an Overlap Class (AAAI 2025 · similarity 0.86)