Class Imbalance: Navigating the AI Frontier with Smart Solutions and Foundational Insights

Latest 50 papers on class imbalance: Nov. 16, 2025

Class imbalance remains one of the most persistent and thorny challenges in machine learning, affecting everything from medical diagnostics to cybersecurity and even space weather forecasting. When one class significantly outnumbers others, models often struggle to learn the characteristics of the minority class, leading to biased predictions and unreliable systems. Fortunately, recent research is brimming with innovative solutions, ranging from novel architectural designs and specialized loss functions to advanced data augmentation techniques and unifying theoretical frameworks.

The Big Idea(s) & Core Innovations

Many recent breakthroughs tackle class imbalance by either refining how models learn from skewed data or by enhancing the data itself. A recurring theme is the move beyond simple rebalancing, emphasizing more nuanced approaches. For instance, in their paper, “Imbalanced Classification through the Lens of Spurious Correlations”, S. Bender et al. from Technische Universität Berlin and BASF SE, argue that class imbalance often leads to spurious correlations, resulting in unreliable classifiers. They introduce Counterfactual Knowledge Distillation (CFKD), a groundbreaking method that leverages counterfactual explanations to explicitly mitigate these issues, demonstrating that traditional loss reweighting is often insufficient.

Similarly, addressing the challenge in multi-view scenarios, Chuanqing Tang et al. from Southwestern University of Finance and Economics propose TMLC in “Trusted Multi-view Learning for Long-tailed Classification”. This framework introduces a group consensus opinion aggregation mechanism, inspired by Social Identity Theory, and an uncertainty-guided pseudo-data generation module to enhance reliability and oversampling effectiveness on long-tailed datasets.

In medical imaging, where class imbalance is particularly critical, multiple papers offer tailored solutions. For example, in “BRIQA: Balanced Reweighting in Image Quality Assessment of Pediatric Brain MRI”, Almsouti et al. (MBZUAI, University of Toronto, Harvard Medical School, Stanford University) introduce BRIQA, which employs tailored model architectures for different MRI artifacts and gradient-based loss reweighting with rotating batching techniques to improve diagnostic reliability. Another significant contribution in this domain comes from Valentyna Starodub and Mantas Lukoševičius (Kaunas University of Technology) in “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”, where they leverage optimized U-Net architectures and weighted binary cross-entropy loss to surpass existing benchmarks for Age-related Macular Degeneration (AMD) lesion detection.

The importance of sophisticated loss functions is further highlighted by Lionel Z. Wang et al. (The Hong Kong Polytechnic University, The University of Hong Kong, Northeastern University) in “SugarTextNet: A Transformer-Based Framework for Detecting Sugar Dating-Related Content on Social Media with Context-Aware Focal Loss”, which proposes Context-Aware Focal Loss (CAFL) to improve minority class detection in sensitive social media content. Similarly, for autonomous driving, Xingcheng Liua et al. from the University of Macau introduce dynamic focal loss in “ROAR: Robust Accident Recognition and Anticipation for Autonomous Driving” to effectively handle class imbalance and noisy data in accident anticipation.

Beyond data and loss functions, theoretical insights are also unifying our understanding of bias. Sushant Mehta’s paper, “When Are Learning Biases Equivalent? A Unifying Framework for Fairness, Robustness, and Distribution Shift”, posits that various bias mechanisms, including class imbalance, can have equivalent effects on model performance, providing a framework for transferring debiasing techniques across different problem types. This ground-breaking theoretical work offers a powerful conceptual lens to address a multitude of AI fairness and robustness challenges.

Under the Hood: Models, Datasets, & Benchmarks

Recent research is not just about new ideas; it’s also about building robust tools and datasets to support these innovations. Here are some key resources and models powering these advancements:

• CABIN Framework: Introduced in “Perceive, Act and Correct: Confidence Is Not Enough for Hyperspectral Classification” by Muzhou Yang et al. (Nanjing University of Aeronautics and Astronautics), this semi-supervised framework for hyperspectral image classification is available on GitHub, leveraging uncertainty-guided dual sampling for improved generalization and label efficiency.
• TMLC (Trusted Multi-view Learning for Long-tailed Classification): The framework with a group consensus opinion aggregation mechanism and uncertainty-guided data generation is open-sourced at GitHub.
• HybridGuard: Proposed by Binayak Kara et al. (National Taiwan University of Science and Technology, Digital University Kerala) in “HybridGuard: Enhancing Minority-Class Intrusion Detection in Dew-Enabled Edge-of-Things Networks”, this ML/DL framework for minority-class intrusion detection utilizes Wasserstein Conditional GANs (WCGAN-GP) and mutual information-based feature selection. The code is available on GitHub, and it’s validated on datasets like UNSW-NB15, CIC-IDS-2017, and IOTID20.
• STARC-9 Dataset: Introduced by Barathi Subramanian et al. (Stanford University, DeepLearning.AI) in “STARC-9: A Large-scale Dataset for Multi-Class Tissue Classification for CRC Histopathology”, this large-scale dataset features 630,000 high-quality tiles across nine colorectal cancer tissue types, addressing morphological diversity and class balance. The dataset and its DeepCluster++ framework are available on Hugging Face and GitHub.
• AnomalyMatch: A semi-supervised active learning framework for anomaly detection combining FixMatch and EfficientNet, presented by Pablo Gómez et al. (European Space Agency, Heidelberg University, University of Groningen) in “AnomalyMatch: Discovering Rare Objects of Interest with Semi-supervised and Active Learning”. Its code is on GitHub, demonstrated on astronomical (GalaxyMNIST) and general (miniImageNet) datasets.
• OCTDL Dataset & CB-ImgComp: For medical diagnostics, “Improving Diagnostic Performance on Small and Imbalanced Datasets Using Class-Based Input Image Composition” by HLALI Azzeddine et al. (Sidi Mohamed Ben Abdellah University) introduces Class-Based Input Image Composition (CB-ImgComp), a novel augmentation strategy for imbalanced medical data, validated on the OCTDL dataset. A related Kaggle dataset is available here.
• RGC Python Package: In “RGC: a radio AGN classifier based on deep learning. I. A semi-supervised model for the VLA images of bent radio AGNs”, M. S. Hossain et al. (Independent University, Bangladesh, Yale University, North Dakota State University) release a Python package with two labeled datasets and a semi-supervised model for classifying bent radio AGNs, utilizing BYOL and E2CNN. The code can be found here.
• N-BEATS and GNN for Anomaly Prediction: Daniel Sorensen et al. (IMEC) in “Unsupervised Anomaly Prediction with N-BEATS and Graph Neural Network in Multi-variate Semiconductor Process Time Series” leverage N-BEATS and GNN for unsupervised anomaly prediction in semiconductor manufacturing. The N-BEATS code is available on GitHub.

These resources underscore a commitment to open science and reproducibility, crucial for advancing research in this complex domain.

Impact & The Road Ahead

The collective impact of this research is profound, touching critical areas from healthcare and cybersecurity to space exploration and autonomous systems. By providing more accurate and reliable models, these advancements can lead to earlier disease detection, stronger defense against cyber threats, better understanding of celestial phenomena, and safer autonomous vehicles.

Looking ahead, the emphasis on explainable AI (XAI) is particularly exciting. Papers like “Interpretable Heart Disease Prediction via a Weighted Ensemble Model: A Large-Scale Study with SHAP and Surrogate Decision Trees” by Md Abrar Hasnat et al. (BRAC University, Linköping University) demonstrate how integrating SHAP and surrogate decision trees can enhance trust in AI predictions in sensitive medical contexts. Similarly, “ConMatFormer: A Multi-attention and Transformer Integrated ConvNext based Deep Learning Model for Enhanced Diabetic Foot Ulcer Classification” by Raihan Ahamed Rifat et al. (Charles Darwin University, University of Liberal Arts Bangladesh, BRAC University, Multimedia University, American International University-Bangladesh) also uses Grad-CAM and LIME for transparency.

The increasing sophistication of hybrid models, combining the strengths of different architectures (e.g., CNNs, attention, Transformers, BiLSTMs), points to a future where highly specialized and efficient solutions can be deployed even in resource-constrained environments, such as the lightweight CNN-Attention-BiLSTM model for arrhythmia detection in wearable ECGs proposed in “A Lightweight CNN-Attention-BiLSTM Architecture for Multi-Class Arrhythmia Classification on Standard and Wearable ECGs”.

Furthermore, the theoretical unification of biases by Sushant Mehta and the data-centric approaches exemplified by “A Data-Centric Approach to Multilingual E-Commerce Product Search: Case Study on Query-Category and Query-Item Relevance” by Yabo Yin et al. promise more principled and robust ways to tackle class imbalance, shifting focus from complex model tweaks to effective data engineering. The journey to truly robust and fair AI systems in the face of imbalanced data is far from over, but these recent breakthroughs provide exciting new directions and powerful tools for the road ahead.