Class Imbalance: Navigating the AI Frontier with Advanced Techniques

Latest 50 papers on class imbalance: Nov. 2, 2025

In the rapidly evolving landscape of AI and machine learning, one persistent challenge frequently surfaces across diverse applications: class imbalance. This occurs when one class significantly outnumbers others in a dataset, often leading to models that perform well on the majority class but fail to adequately recognize or predict the minority – which can be the most critical, like detecting a rare disease or an anomalous event. Recent research, as highlighted in a collection of groundbreaking papers, reveals a surge in innovative solutions addressing this fundamental problem, pushing the boundaries of what AI can achieve in complex, real-world scenarios.

The Big Idea(s) & Core Innovations

The core challenge in class-imbalanced datasets lies in training models that are both accurate and robust across all classes, especially the under-represented ones. Researchers are tackling this by refining model architectures, developing sophisticated loss functions, and employing novel data augmentation strategies.

In medical imaging, where class imbalance is rampant (e.g., rare lesions or diseases), the focus is on robust segmentation and classification. For instance, Valentyna Starodub and Mantas Lukoševičius from Kaunas University of Technology, in their paper “Surpassing state of the art on AMD area estimation from RGB fundus images through careful selection of U-Net architectures and loss functions for class imbalance”, demonstrate how optimized U-Net architectures combined with a weighted binary cross-entropy loss significantly improve AMD lesion detection. Similarly, Almsouti et al. from MBZUAI, University of Toronto, and others, in “BRIQA: Balanced Reweighting in Image Quality Assessment of Pediatric Brain MRI”, introduce tailored models and gradient-based reweighting, alongside a rotating batching technique, to classify MRI artifact severity, notably in low-field systems.

Another innovative approach to medical imaging is presented by E. Gad et al. from the University of Cairo and Medical AI Research Lab, in “Advancing Brain Tumor Segmentation via Attention-based 3D U-Net Architecture and Digital Image Processing”. They integrate attention mechanisms into 3D U-Nets and use digital image processing to balance class distribution, achieving high accuracy in brain tumor segmentation. Meanwhile, Asha et al., in “SG-CLDFF: A Novel Framework for Automated White Blood Cell Classification and Segmentation”, leverage saliency detection and cross-layer deep feature fusion with class-aware weighted loss functions to tackle WBC classification.

The critical role of data augmentation is highlighted in several papers. Vaishnavi Visweswaraiah et al. from Harrisburg University of Science & Technology and Wright State University, in “Handling Extreme Class Imbalance: Using GANs in Data Augmentation for Suicide Prediction”, show that GANs can generate synthetic data to significantly improve the detection of rare suicide attempt cases. Extending this, Sasan Farhadi et al. from Politecnico di Torino and ETH Zürich, in “Addressing data scarcity in structural health monitoring through generative augmentation”, introduce STFTSynth, a WGAN-GP-based model for generating realistic spectrograms to detect rare events like wire breakage in structural health monitoring. In drug discovery, Xin Wang et al. from Yale University and University of Oregon, in “Scaffold-Aware Generative Augmentation and Reranking for Enhanced Virtual Screening”, propose ScaffAug, a framework that uses graph diffusion models for scaffold-aware augmentation to tackle class and structural imbalances in virtual screening, leading to the discovery of diverse active compounds.

Addressing class imbalance isn’t just about data; it’s also about tailored evaluation metrics. Pierangelo Lombardo et al. from Eutelsat and Reply, in “Cost-Sensitive Evaluation for Binary Classifiers”, introduce Weighted Accuracy (WA) as a new metric that aligns with minimizing Total Classification Cost (TCC), offering a robust way to compare models in cost-sensitive, imbalanced scenarios.

Under the Hood: Models, Datasets, & Benchmarks

These innovations are often built upon advancements in core models and specialized datasets, many of which are now publicly available, fostering reproducibility and further research. Here’s a glimpse:

• U-Net Architectures & Weighted Binary Cross-Entropy Loss: Optimized U-Nets for AMD lesion detection, achieving superior performance on the ADAM challenge. Code available at https://github.com/vlntn-starodub/AMD-lesion-segmentation.
• BRIQA Framework & Rotating Batching: Tailored models and gradient-based reweighting for pediatric brain MRI artifact classification. Public code via https://github.com/BioMedIA.
• Generative Adversarial Networks (GANs): Used in “Pulsar Detection with Deep Learning” by Manideep Pendyala (IISER Bhopal) to generate synthetic data, improving pulsar classification accuracy to 94%. Code available at https://github.com/manideep-pendyala/pulsar-detection.
• AnomalyMatch (FixMatch + EfficientNet): A semi-supervised active learning framework for anomaly detection, achieving high AUROC/AUPRC with minimal labels. Integrated into ESA Datalabs. Code at https://github.com/esa/AnomalyMatch.
• ConMatFormer: A hybrid model for Diabetic Foot Ulcer (DFU) classification, combining ConvNeXt, attention mechanisms, and transformers. Addresses class imbalance with data augmentation and explains predictions using Grad-CAM/LIME. See details in “ConMatFormer: A Multi-attention and Transformer Integrated ConvNext based Deep Learning Model for Enhanced Diabetic Foot Ulcer Classification”.
• DenseNet121 with Hybrid Loss: Utilized in “Robust Atypical Mitosis Classification with DenseNet121: Stain-Aware Augmentation and Hybrid Loss for Domain Generalization” by Dukre et al. (MBZUAI), employing stain-aware augmentation and a combined class-weighted binary cross-entropy and focal loss for robust atypical mitosis classification.
• RGC Python Package: Includes two labeled datasets and a semi-supervised model using BYOL and E2CNN for classifying bent radio AGNs. Public code at https://github.com/inigoval/fixmatch from M. S. Hossain et al. (Independent University, Bangladesh).
• Weighted Accuracy (WA): A new metric for binary classifiers, presented in “Cost-Sensitive Evaluation for Binary Classifiers”. Code at https://github.com/plombardML/weighted-accuracy.
• STFTSynth (WGAN-GP): A generative model for synthetic spectrograms in structural health monitoring, available at https://github.com/sasanfarhadi/STFTSynth.
• HACO Framework: Proposed by Daniel Sungho Jung and Kyoung Mu Lee from Seoul National University in “Learning Dense Hand Contact Estimation from Imbalanced Data”, using Balanced Contact Sampling (BCS) and Vertex-Level Class-Balanced (VCB) loss for dense hand contact estimation. Code: https://github.com/dqj5182/HACO_RELEASE.
• EndoCIL Framework: A class-incremental learning framework for endoscopic images, featuring MDBR, PRCBL, and CFG. Detailed in “EndoCIL: A Class-Incremental Learning Framework for Endoscopic Image Classification”.
• xLSTM: A parameter-efficient architecture by Noor Islam S. Mohammad (New York University) for toxic comment classification, integrating cosine-similarity gating. Code at https://github.com/csislam/xLSTM-Project.
• YOLO-World & Spatio-Temporal Deep Learning: Integrated for human-centric anomaly detection in surveillance videos, as presented in “Human-Centric Anomaly Detection in Surveillance Videos Using YOLO-World and Spatio-Temporal Deep Learning” by T. Li et al. (Chinese Academy of Sciences and University of Science and Technology of China).
• DeBERTa-KC: A transformer-based classifier by Jindi Wang et al. (Durham University) for knowledge construction in online learning discourse, utilizing Focal Loss, Label Smoothing, and R-Drop regularization. Paper: “DeBERTa-KC: A Transformer-Based Classifier for Knowledge Construction in Online Learning Discourse”.
• Dr.LLM: A dynamic layer routing framework for LLMs by Ahmed Heakl et al. (Parameter Lab, MBZUAI, NAVER AI Lab), using Monte Carlo Tree Search (MCTS) and Focal Loss with class-balancing for efficiency. Code: https://github.com/parameterlab/dr-llm.
• DEF-YOLO: Modified YOLOv8 with deformable convolution for concealed weapon detection in thermal imaging. Introduces the TICW dataset. Paper: “DEF-YOLO: Leveraging YOLO for Concealed Weapon Detection in Thermal Imaging”.
• GLOFNet Dataset: A multimodal dataset for Glacial Lake Outburst Flood (GLOF) monitoring, integrating Sentinel-2, ITS_LIVE, and MODIS LST data. Paper: “GLOFNet – A Multimodal Dataset for GLOF Monitoring and Prediction”.
• Cyc-Attack: A gradient-based adversarial attack for tropical cyclone trajectory prediction, using a differentiable surrogate model and skewness-aware loss for class imbalance. Code: https://github.com/dengy0111/Cyc-Attack.

Impact & The Road Ahead

The collective insights from these papers represent a significant leap forward in tackling class imbalance across diverse fields, from medicine and astronomy to cybersecurity and e-commerce. The development of tailored loss functions, sophisticated data augmentation techniques (especially with generative models), and new evaluation metrics are enabling more robust and reliable AI systems.

For instance, the ability to accurately detect rare medical conditions, predict critical environmental events, or identify subtle anomalies in system logs has profound real-world implications, leading to better diagnostic tools, enhanced safety, and improved resource allocation. The work on SMOTE and Mirrors by Georgi Ganev et al. (SAS, UCL, UC Riverside) in “SMOTE and Mirrors: Exposing Privacy Leakage from Synthetic Minority Oversampling” also serves as a crucial reminder that while synthetic data helps address imbalance, privacy considerations must be paramount, opening avenues for future research into privacy-preserving data augmentation.

The emphasis on open-source code and standardized benchmarks, seen in papers like “Long-tailed Species Recognition in the NACTI Wildlife Dataset” by Tabak, M. A. et al. (LILA Science), fosters a collaborative environment for researchers to build upon these advancements. The future of AI in dealing with class imbalance will likely see further integration of multimodal data, more dynamic and adaptive learning frameworks, and an even stronger focus on explainability and ethical considerations, ensuring that the power of AI is harnessed responsibly for the benefit of all.