Class Imbalance No More: Recent Breakthroughs in AI/ML for Robust and Fair Systems

Latest 50 papers on class imbalance: Sep. 1, 2025

Class imbalance is a pervasive and often insidious challenge in machine learning, where the unequal distribution of samples across categories can lead to models that perform poorly on under-represented classes, despite achieving high overall accuracy. This isn’t just an academic problem; it impacts critical real-world applications from medical diagnostics to fraud detection and cybersecurity. Fortunately, recent research has unveiled a host of innovative solutions, pushing the boundaries of what’s possible in building more robust, fair, and effective AI systems.

The Big Idea(s) & Core Innovations

The papers summarized here tackle class imbalance from diverse angles, often proposing multi-pronged approaches that combine novel architectures, loss functions, and data augmentation strategies. A recurring theme is the move beyond simple oversampling to more sophisticated, context-aware methods.

In the realm of long-tailed recognition and visual systems, a significant advancement comes from X. Wei and Haibo Ye (National University of Defense Technology, Nanjing University of Aeronautics and Astronautics) with their paper, “Divide, Weight, and Route: Difficulty-Aware Optimization with Dynamic Expert Fusion for Long-tailed Recognition”. They introduce DQRoute, a framework that recognizes class frequency isn’t always a proxy for learning difficulty. Instead, it dynamically routes and reweights ‘experts’ based on prediction uncertainty and accuracy, significantly boosting performance on rare, critical tail classes. Similarly, W. Kularatne et al. (University of Oulu) address object detection in UAV surveillance in “Exponentially Weighted Instance-Aware Repeat Factor Sampling for Long-Tailed Object Detection Model Training in Unmanned Aerial Vehicles Surveillance Scenarios” by proposing E-IRFS, an instance-aware repeat factor sampling method that makes training more efficient for detecting rare objects under resource constraints.

For medical imaging and diagnostics, where class imbalance is particularly acute (e.g., rare diseases), several papers offer compelling solutions. Dennis Slobodzian et al. (University of Southern Maine, University of Maine) in “Deep Learning Framework for Early Detection of Pancreatic Cancer Using Multi-Modal Medical Imaging Analysis” show that optimized transfer learning and class-weighted training on dual-modality imaging significantly improves early PDAC detection, achieving over 90% accuracy on limited datasets. “Automated Classification of Normal and Atypical Mitotic Figures Using ConvNeXt V2: MIDOG 2025 Track 2” by Yosuke Yamagishi and Shouhei Hanaoka (The University of Tokyo) demonstrates how center cropping and cross-validation with ConvNeXt V2 improve robustness against class imbalance and domain heterogeneity in mitotic figure classification. Md. Marufa et al. (Rajshahi University of Engineering & Technology, Bangladesh) push the boundaries of Acute Lymphoblastic Leukemia (ALL) diagnosis in “Deep Learning with Self-Attention and Enhanced Preprocessing for Precise Diagnosis of Acute Lymphoblastic Leukemia from Bone Marrow Smears in Hemato-Oncology” by integrating Multi-Head Self-Attention with Focal Loss, achieving a remarkable 99.25% accuracy. Lishi Zuo et al. (The Hong Kong Polytechnic University, Hong Kong SAR) tackle the nuanced ‘class-feature bias’ in “Class Unbiasing for Generalization in Medical Diagnosis”, proposing a class-unbiased model (Cls-unbias) with class-wise inequality loss and Group Distributionally Robust Optimization (G-DRO) to improve generalization by leveraging class-shared features.

Fraud detection and cybersecurity also see innovative solutions. Yudan Song et al. (Guangxi Normal University, Beihang University, Tianjin University) introduce MimbFD in “Mitigating Message Imbalance in Fraud Detection with Dual-View Graph Representation Learning” to address message imbalance in Graph Neural Networks (GNNs), using a dual-view approach to enhance representation learning against class and topological imbalances. Xinrui Li et al. (Chongqing University, The Hong Kong University of Science and Technology (Guangzhou), Nanjing University) propose GraphFedMIG in “GraphFedMIG: Tackling Class Imbalance in Federated Graph Learning via Mutual Information-Guided Generation”, a federated generative data augmentation task that uses mutual information-guided mechanisms to prioritize minority-class patterns in federated graph learning, significantly improving performance on rare fraud instances.

Beyond these specific domains, the fundamental idea of synthetic data generation for imbalance is gaining traction. Mika Leo Hube et al. (Polytechnic University of Catalonia, i2Cat Foundation) in “Set Transformer Architectures and Synthetic Data Generation for Flow-Guided Nanoscale Localization” use CGAN, WGAN, and CVAE to overcome data scarcity and class imbalance for nanoscale localization. Similarly, Jin Yang (Sichuan University, China) introduces MixGAN in “MixGAN: A Hybrid Semi-Supervised and Generative Approach for DDoS Detection in Cloud-Integrated IoT Networks”, combining semi-supervised learning with CTGAN-based generative augmentation for DDoS detection. A. T. Wasi et al. (Information Sciences Institute, University of Southern California) apply GFlowNets with Variational Graph Autoencoders in “GFlowNets for Learning Better Drug-Drug Interaction Representations” to generate synthetic DDI samples, boosting predictions for rare drug interactions.

Under the Hood: Models, Datasets, & Benchmarks

These advancements are often powered by clever combinations of established and emerging techniques, along with new datasets and rigorous benchmarking:

• Deep Learning Architectures: The ubiquitous ResNet appears in the pancreatic cancer detection paper, where it’s optimized with frozen pre-trained layers. ConvNeXt V2 is highlighted for its balance of efficiency and capacity in mitotic figure classification. For semantic segmentation, the novel DBIF-AUNet (from Ruixiang Tang et al., Yunnan University) significantly outperforms U-Net++ and Swin-UNet in pleural effusion detection by leveraging dual-domain feature disentanglement and branch interaction attention fusion. EfficientNet-B3 (Faisal Ahmed, Embry-Riddle Aeronautical University) proves superior for Acute Lymphoblastic Leukemia classification, demonstrating the power of transfer learning and data augmentation. In fraud detection, Graph Neural Networks (GNNs) are central, with dual-view approaches like MimbFD and generative methods like GraphFedMIG pushing their capabilities.

• Generative Models & Data Augmentation: Techniques like CTGAN, WGAN, WGAN-GP, and CVAE are employed for conditional tabular synthesis and synthetic data generation, particularly in network security (MixGAN) and nanoscale localization. SMOTETomek is integrated client-side in federated learning (Rodrigo Ronner Tertulino da Silva, Software Engineering and Automation Research Laboratory) to address class imbalance in clinical data, demonstrating its effectiveness in a privacy-preserving context. Feature-Space Oversampling is also highlighted as an effective strategy for SAR ship classification (Awais Khan et al., University of Central Michigan).

• Loss Functions & Optimization: Focal Loss is a recurring solution for class imbalance, seen in ALL diagnosis and multi-task speech emotion recognition (Honghong Wang et al., Beijing Fosafer Information Technology Co., Ltd., China) where it’s combined with a novel Sample Weighted Focal Contrastive (SWFC) loss. The medical diagnosis paper by Lishi Zuo et al. introduces a class-wise inequality loss combined with Group Distributionally Robust Optimization (G-DRO).

• Novel Frameworks & Mechanisms: Beyond architectures, several papers introduce comprehensive frameworks: DQRoute for dynamic expert fusion, MixGAN for hybrid semi-supervised learning, MimbFD for dual-view graph representation learning, E-CaTCH (Ahmad Mousavia et al., American University) for event-centric cross-modal attention in misinformation detection, and MDPR (Yongju Jia et al., Shandong University) for LLM-empowered dynamic prompt routing in Vision-Language Models. The SubROC framework (Tom Siegl et al., University of Rostock) provides an open-source tool (https://doi.org/10.5281/zenodo/16952343) for discovering exceptional subgroup performance based on AUC measures, incorporating class imbalance control. The concept of Kolmogorov-Arnold Networks (KAN) is also emerging, with KARMA (https://github.com/faeyelab/) using it for efficient structural defect segmentation and KACQ-DCNN (Md Abrar Jahin et al., University of Southern California) integrating it into classical-quantum hybrid networks for heart disease detection.

• Datasets & Code: Researchers are actively contributing new datasets and making code publicly available. Examples include a high-quality CT scan dataset for sarcopenia detection (Bhardwaj et al., Freeman Hospital, Newcastle upon Tyne, UK), the BD-TypoSAT dataset for post-disaster damage assessment (Yiming Xiao and Ali Mostafavi, Texas A&M University), and various GitHub repositories for models like E-IRFS (https://github.com/futurians/E-IRFS), SDGNN (https://github.com/mingyue15694/SGDNN/tree/main), MixGAN (https://github.com/0xCavaliers/MixGAN), CRoC (https://github.com/XsLangley/CRoC_ECAI2025), and GraphFedMIG (https://github.com/NovaFoxjet/GraphFedMIG).

Impact & The Road Ahead

These research efforts collectively represent a significant leap forward in tackling class imbalance, an issue fundamental to deploying reliable AI systems. The practical implications are profound: more accurate early cancer detection, robust fraud prevention, dependable anomaly detection in cyber-physical systems, and fairer credit assessment. The shift towards difficulty-aware, context-sensitive, and generative approaches marks a maturation in how we address this challenge.

The road ahead involves further integration of causal inference (as seen in “Counterfactual Reward Model Training for Bias Mitigation in Multimodal Reinforcement Learning” by Sheryl Mathew and N Harshit, Vellore Institute of Technology) and explainable AI (XAI) techniques. Papers like “Explainable AI (XAI) for Arrhythmia detection from electrocardiograms” (Joschka Beck and Arlene John, University of Twente) and “Explainable AI Technique in Lung Cancer Detection Using Convolutional Neural Networks” (Nishan Rai et al., Kathford International College of Engineering and Management) highlight the critical need for interpretability alongside performance, especially in sensitive domains like healthcare. Furthermore, the increasing adoption of federated learning requires robust solutions for imbalance in distributed settings, a challenge GraphFedMIG successfully addresses. As AI continues to permeate every facet of our lives, the ability to build models that are not only powerful but also fair, transparent, and robust to real-world data imperfections will be paramount. This wave of research brings us closer to that reality, promising a future where AI’s benefits are equitably distributed and reliably delivered.