Class Imbalance: Navigating the Edge Cases of AI/ML

Latest 27 papers on class imbalance: Oct. 31, 2026

The world of AI/ML is often a battle against the imbalanced. From detecting rare diseases to spotting fraudulent transactions, real-world data rarely comes perfectly packaged. This fundamental challenge, where one class significantly outnumbers another, poses a substantial hurdle to building robust, fair, and effective machine learning models. Recently, a flurry of innovative research has tackled this problem head-on, offering novel solutions that promise to enhance model performance, explainability, and ethical deployment in critical applications.

The Big Idea(s) & Core Innovations

Many recent breakthroughs converge on a central theme: treating class imbalance not just as a data quantity problem, but as a deeper issue intertwined with data complexity, model interpretability, and robust generalization. The paper, “CBE: A Complexity-adaptive Boundary-aware Ensemble Method for Imbalanced Data” by Xun Wei from Jiangxi University of Science and Technology, introduces an Overall Complexity Measure (OCM) to categorize datasets into ‘Easy,’ ‘Median,’ and ‘Hard’ based on local overlap, boundary geometry, and global separability. This innovative approach allows for Complexity-adaptive Distance-Weighted Sampling (CDWS), which intelligently prioritizes boundary examples only when the decision boundary is clear, avoiding overfitting noise in highly overlapping datasets. This signals a shift from indiscriminate sampling to informed, context-aware strategies.

Building on the need for adaptive approaches, “Inference-Time Prior Adaptation for Imbalanced Data Streams under Concept Drift” by Tobias Callies, Eirini Ntoutsi, and Arthur Zimek demonstrates that adapting to changing class distributions in data streams doesn’t always require complex memory restructuring or synthetic data generation. Their PAI-KNN method, a sliding-window k-nearest-neighbor classifier, dynamically adjusts its inference rule using a statistically consistent label history. This highlights that for dynamic imbalances and concept drift, smarter inference can be as effective as complex data manipulation.

In critical domains like healthcare, where the cost of a false negative is high, managing imbalance becomes paramount. “When uncertainty quantification fails: structural vulnerabilities and corrections for selective prediction under clinical class imbalance” by Walter Endrizzi and colleagues from Fondazione Bruno Kessler and other institutions, exposes how standard uncertainty quantification can systematically defer minority patients. They propose a threshold-aware ambiguity score (Hτ) and Class-Conditioned Rejection (CCR), which ranks and rejects patients separately within each predicted clinical trajectory. This prevents global uncertainty metrics from masking minority exclusion and ensures equitable risk assessment. This work underscores that addressing imbalance requires a holistic approach that considers fairness and ethical implications from the outset.

Furthermore, the realm of self-supervised learning and multimodal data is seeing significant advancements. “TransECG-SSL: Hybrid Self-Supervised Representation Learning for ECG Arrhythmia Classification” by Ahtisham Ayyub et al. from Kingston University and the University of Edinburgh, showcases a Transformer-based framework that combines masked ECG reconstruction and contrastive learning. This hybrid approach learns transferable representations from unlabeled ECG data, demonstrating strong performance in cross-dataset arrhythmia detection, even with severe minority-class imbalance. Similarly, “Internal Discrepancy for Contrastive Learning: Generalization Bounds and Downstream Behaviors” by Jialiang Wang and colleagues from Harbin Institute of Technology provides a theoretical foundation, showing how internal discrepancy in contrastive learning can mitigate class imbalance and lead to more structured, optimal representations.

For practical deployment, methods that generate synthetic data are gaining traction. “Enhancing Deep Learning Model Performance through Data Generation Strategies: Classification of Olive Tree Diseases” by Hicham El Akhal et al. from Moulay Ismail University of Meknes, evaluates traditional augmentation, MixUp, CutMix, and DCGAN-generated images for olive disease classification. They find that targeted generation, especially with DCGANs, significantly improves minority-class representation and overall performance. In a groundbreaking move for medical imaging, “Pos4CHD: A Positive-Free Mask-Guided Counterfactual Synthesis Framework for Fetal Congenital Heart Defect Analysis” by Wenjie Xuan et al. from Wuhan University, introduces Pos4CHD, which synthesizes pseudo-positive fetal echocardiograms for congenital heart defect analysis without real positive training samples. This is achieved by manipulating anatomical masks on normal images and using a Conditional Denoising Diffusion Operator with histogram regularization to improve ultrasound fidelity. This opens new avenues for rare disease detection where positive samples are extremely scarce.

Under the Hood: Models, Datasets, & Benchmarks

Recent research heavily leverages existing benchmarks while also introducing new, more challenging ones to better evaluate real-world performance under imbalance:

• Complexity-Adaptive Ensemble Methods: The CBE framework (from “CBE: A Complexity-adaptive Boundary-aware Ensemble Method for Imbalanced Data”) was validated on 24 diverse UCI and KEEL datasets, demonstrating its ability to adapt sampling strategies based on dataset complexity. Code is available at https://github.com/weixun2018/CBE.
• ECG Arrhythmia Detection: “Enhancing Atrial Fibrillation Detection Through Hyperparameter Optimization” used the MIT-BIH Arrhythmia and Long-Term AF databases. Similarly, “TransECG-SSL: Hybrid Self-Supervised Representation Learning for ECG Arrhythmia Classification” used MIT-BIH, INCART, and SVDB databases for cross-dataset generalization with a Transformer encoder. Code for TransECG-SSL is at https://github.com/ahtisham-86/TransECG-SSL.
• Fraud Detection: In “Explainable and Adaptive Machine Learning for Fraud Risk Detection in Bangladesh’s Digital Banking Ecosystem”, experiments were conducted on the synthetic PaySim benchmark, using models like Clean XGBoost and LightGBM. “RGH-GNN: Risk-Gated Heterogeneous Graph Neural Network for Financial Fraud Detection” utilized the IEEE-CIS Fraud Detection dataset, employing LightGBM for transaction-risk scoring and various GNNs (GraphSAGE, GIN, GAT, etc.).
• Medical Image Segmentation & Classification: A variety of datasets and models are highlighted. For Alzheimer’s, “How Robust Are CNNs and Vision Transformers for Alzheimer’s MRI Classification Across Datasets?” benchmarked ResNet-50, ReXNet-150, Swin Transformer, ViT-Base, and ViT-Large across multiple MRI datasets. “EigenGradCAM with AugSmooth: Enhancing Explainability and Precision in Multi-Stage Alzheimer’s Disease Detection” used a custom CNN with WGAN-GP synthetic augmentation on a Kaggle Alzheimer’s MRI dataset. O2D-MedNet (from “O2D-MedNet: An Optimization-Aware Dual-Domain Framework for Robust Medical Image Segmentation”) was evaluated on six medical imaging datasets (Oral, Kvasir-SEG, COVID-19 CT, MLUA, CT2US, BUSI) using CNN and visual state-space model (VSSM) backbones. For surgical segmentation, “Rethinking Data Imbalance in Class Incremental Surgical Instrument Segmentation” introduced SurgCSS on a benchmark combining EndoVis 2017/2018, CholecSeg8k, and SAR-RARP50, available at https://github.com/Zzsf11/SurgCSS.
• Predictive Maintenance: “Pneuma: A Large-Scale Real-World Time Series Benchmark for Commercial Aircraft Predictive Maintenance” introduced Pneuma, a massive dataset of 920 million QAR time steps from commercial aircraft, benchmarking over 25 models, including time-series foundation models like Toto-2.0 and Chronos-2. While the dataset is publicly available, a direct download link is not in the summary.
• Diabetes Diagnosis: “Enhancing Type 2 Diabetes Diagnosis with Evolutionary Algorithms and Machine Learning” used the classic Pima Indians Diabetes Database, combining SMOTE-ENN with a Bat Optimization Algorithm for feature selection and Random Forest tuning.
• Functional Data Augmentation: “Synthetic Data Augmentation for Functional Data: Wasserstein Risk Bounds and Diffusion-Based Conditional Generation” provided empirical validation on Spanish temperature curves, grape hyperspectral spectra, and gait biomechanics data, introducing a Conditional Denoising Diffusion Operator.
• Online Continual Learning: “CARES: Coverage-Aware Replay Exemplar Selection for Continual Learning” evaluated its CARES method on Split CIFAR-10, Split CIFAR-100, and Split Tiny-ImageNet, utilizing frozen pretrained encoders like Inception-V3, ResNet-50, and CLIP ViT-B/32.
• Performance Estimation: “Estimating Performance for Safe Deployment of Machine Learning Models” introduced MOVAL, a framework for label-free performance estimation with class-specific calibration for classification and 2D/3D segmentation, available as a PyPI package: https://pypi.org/project/moval/.
• Depth Sensor Simulation: “PRISM: Learning Realistic Depth via Physics-Grounded Noise Disentanglement with Semantic-Geometric Collaboration” uses 3D visual foundation model priors like MoGe-2 and Depth Anything 3 for synthesizing realistic depth sensor noise, evaluated on ByteCameraDepth and NYU-Depth-v2. Project page: https://shawnleung-sii.github.io/prism-icml2026/.

Impact & The Road Ahead

The collective impact of this research is profound. By moving beyond simplistic rebalancing to more sophisticated, context-aware, and explainable approaches, these advancements are paving the way for more robust and trustworthy AI systems, particularly in sensitive areas like healthcare and finance. The emphasis on explainability (SHAP, EigenGradCAM), privacy (Fed-IGTD, DP-FedAvg), and fairness (class-conditioned rejection) signals a maturing field keenly aware of its societal responsibilities.

Looking ahead, several themes emerge. The development of physically grounded synthetic data generation (Pos4CHD, PRISM) promises to unlock AI applications in data-scarce domains. Adaptive learning strategies that dynamically adjust to data complexity and drift (CBE, PAI-KNN) will be crucial for real-time, online systems. Furthermore, the robust benchmarking of Vision Transformers (ViT, Swin) and the insights into self-supervised learning (TransECG-SSL) will continue to drive foundational model development. The challenge of class imbalance remains, but these innovations demonstrate that a multi-faceted, interdisciplinary approach, combining theoretical rigor with practical engineering, is steadily bringing us closer to AI that performs exceptionally, transparently, and equitably in the most challenging real-world scenarios.

Please leave this field empty

Hi there 👋

Get a roundup of the latest AI paper digests in a quick, clean weekly email.

Check your inbox or spam folder to confirm your subscription.