Class Imbalance No More: Recent Breakthroughs in Tackling Skewed Data in AI/ML

Latest 98 papers on class imbalance: Aug. 17, 2025

Class imbalance remains one of the most pervasive and insidious challenges in real-world AI and Machine Learning applications. From detecting rare diseases to identifying fraudulent transactions or critical system anomalies, datasets often feature a vast majority of ‘normal’ cases and a tiny minority of ‘interesting’ or ‘critical’ ones. This inherent skew can lead models to become biased towards the dominant class, resulting in misleadingly high accuracy but abysmal performance on the very events we care most about. Fortunately, recent research is pushing the boundaries of how we tackle this problem, moving beyond simple resampling to more sophisticated, integrated, and even quantum-inspired solutions.### The Big Idea(s) & Core Innovationsthe heart of these advancements is a multifaceted approach that extends beyond traditional statistical remedies. Researchers are leveraging advanced generative models, clever architectural designs, and even theoretical re-evaluations of what “balanced” truly means. For instance, in medical imaging, the challenge is particularly acute. The paper VasoMIM: Vascular Anatomy-Aware Masked Image Modeling for Vessel Segmentation from researchers at the State Key Laboratory of Multimodal Artificial Intelligence Systems, Chinese Academy of Sciences, introduces VasoMIM, a masked image modeling framework that embeds vascular anatomy into pre-training, enhancing representations and addressing class imbalance in X-ray angiograms. Similarly, in lung cancer detection, the Multi-Attention Stacked Ensemble for Lung Cancer Detection in CT Scans by Uzzal Saha and Surya Prakash from the Indian Institute of Technology Indore, uses a dual-level attention mechanism and Dynamic Focal Loss to significantly reduce error rates and improve robustness against class imbalance.power of synthetic data generation is a recurring theme. Enhancing Glass Defect Detection with Diffusion Models by Sajjad Rezvani Boroujeni et al. from Bowling Green State University and Actual Reality Technologies, applies Denoising Diffusion Probabilistic Models (DDPMs) to create realistic defective glass images, dramatically boosting recall for rare defects without false positives. This resonates with the innovation in drug discovery, where GFlowNets for Learning Better Drug-Drug Interaction Representations from A. T. Wasi et al. at the Information Sciences Institute, University of Southern California, combines GFlowNets with Variational Graph Autoencoders (VGAE) to generate synthetic DDI samples, ensuring better predictions for rare drug interactions. In the realm of cyber threat intelligence, SynthCTI: LLM-Driven Synthetic CTI Generation to enhance MITRE Technique Mapping by Álvaro Ruiz-Ródenas et al. demonstrates how LLMs can generate high-quality synthetic CTI sentences for underrepresented MITRE ATT&CK techniques, boosting classification performance.generation, specialized model architectures and learning paradigms are emerging. GraphFedMIG: Tackling Class Imbalance in Federated Graph Learning via Mutual Information-Guided Generation by Xinrui Li et al. from Chongqing University, redefines federated graph learning as a generative data augmentation task, ensuring minority class patterns are preserved. For object detection, DyCAF-Net: Dynamic Class-Aware Fusion Network integrates class-aware feature fusion, enabling accurate detection even in complex, imbalanced environments. A theoretical dive into the problem, Class Imbalance in Anomaly Detection: Learning from an Exactly Solvable Model by F.S. Pezzicoli et al. challenges the conventional wisdom that perfectly balanced training sets are always optimal, revealing how intrinsic imbalance and data abundance influence performance in anomaly detection.### Under the Hood: Models, Datasets, & Benchmarksresearch heavily relies on and contributes to diverse models, datasets, and benchmarks, showcasing the broad applicability of these solutions:VasoMIM: Utilizes anatomy-guided masked image modeling for X-ray angiograms, achieving SOTA on three benchmarks.Adapting SAM via Cross-Entropy Masking: Improves the Segment Anything Model (SAM) for remote sensing change detection on datasets like S2Looking, achieving 2.5% F1-score gain.GraphFedMIG: A novel FGL paradigm that uses mutual information-guided generative data augmentation across multiple real-world datasets (code available).Understanding Textual Emotion Through Emoji Prediction: Evaluates BERT, CNN, and Transformer models on the TweetEval dataset, highlighting focal loss for rare emoji classes (code available).A Robust Pipeline for Differentially Private Federated Learning on Imbalanced Clinical Data: Validates a multi-stage pipeline with client-side SMOTETomek and FedProx on the stroke prediction dataset from Kaggle.MOTGNN: An interpretable Graph Neural Network framework for multi-omics disease classification, robust to severe class imbalance on datasets like TCGA.Class Unbiasing for Generalization in Medical Diagnosis: Introduces ‘class-feature bias’ and a class-unbiased model (Cls-unbias) with group distributionally robust optimization (code available).GFlowNets for Learning Better Drug-Drug Interaction Representations: Integrates GFlowNets with VGAE for synthetic DDI sample generation (code available).Transfer Learning with EfficientNet for Accurate Leukemia Cell Classification: Employs EfficientNet variants with data augmentation on the C-NMC Challenge dataset (code available).DamageCAT: Introduces the BD-TypoSAT dataset for typology-based post-disaster building damage assessment, using a hierarchical U-Net transformer (code available).Feature-Space Oversampling for Addressing Class Imbalance in SAR Ship Classification: Uses lightweight CNN with feature-space oversampling (code available).A Semantic Segmentation Algorithm for Pleural Effusion Based on DBIF-AUNet: Introduces DBIF-AUNet with hierarchical adaptive hybrid loss for class imbalance in CT images.SMOGAN: Synthetic Minority Oversampling with GAN Refinement for Imbalanced Regression: Uses DistGAN to refine synthetic samples for imbalanced regression tasks, outperforming on 23 benchmark datasets.F2PASeg: Feature Fusion for Pituitary Anatomy Segmentation in Endoscopic Surgery: Proposes F2PASeg and a large-scale Pituitary Anatomy Segmentation (PAS) dataset (code available).An Explainable Machine Learning Framework for Railway Predictive Maintenance: Leverages the MetroPT dataset with online learning and XAI for real-time fault detection (code available).ALScope: A Unified Toolkit for Deep Active Learning: A platform supporting 21 DAL algorithms across 10 datasets, customizable for class imbalance (code available).Multi-Stage Knowledge-Distilled VGAE and GAT for Robust Controller-Area-Network Intrusion Detection: Uses VGAE and KD-GAT on six public CAN intrusion datasets for lightweight, accurate detection.Proto-EVFL: Enhanced Vertical Federated Learning via Dual Prototype with Extremely Unaligned Data: Addresses data unalignment in VFL with dual prototypes.An Enhanced Focal Loss Function to Mitigate Class Imbalance in Auto Insurance Fraud Detection with Explainable AI: Introduces a customized multistage focal loss function for fraud detection.Understanding the Essence: Delving into Annotator Prototype Learning for Multi-Class Annotation Aggregation: Proposes PTBCC, a prototype learning-driven Bayesian classifier for multi-class annotation aggregation (code available).CLIMD: A Curriculum Learning Framework for Imbalanced Multimodal Diagnosis: A plug-and-play framework for multimodal medical diagnosis, avoiding data augmentation via curriculum learning (code available).GeHirNet: A Gender-Aware Hierarchical Model for Voice Pathology Classification: A two-stage framework leveraging gender-specific patterns and advanced data augmentation (code available).Multi-VQC: A Novel QML Approach for Enhancing Healthcare Classification: Introduces a hybrid quantum-classical ML framework for imbalanced medical datasets (code available).A Conditional GAN for Tabular Data Generation with Probabilistic Sampling of Latent Subspaces: Generates high-quality, balanced synthetic tabular data.Diffusion-Based User-Guided Data Augmentation for Coronary Stenosis Detection: Uses diffusion models and ControlNet for synthetic coronary angiograms with controlled stenosis severity (code available).CXR-CML: Improved zero-shot classification of long-tailed multi-label diseases in Chest X-Rays: Models latent space using GMM and Student-t distribution for rare diseases.Robust Five-Class and binary Diabetic Retinopathy Classification Using Transfer Learning and Data Augmentation: Utilizes transfer learning and class-balanced data augmentation on the APTOS 2019 dataset (code available).Kolmogorov Arnold Networks (KANs) for Imbalanced Data – An Empirical Perspective: Empirically evaluates KANs on ten benchmark datasets, showing promise on raw imbalanced data but limitations with resampling.### Impact & The Road Aheadrecent breakthroughs signal a significant shift in how the AI/ML community addresses class imbalance. The emerging strategies emphasize not just balancing numbers but creating more meaningful and clinically/practically relevant representations of minority classes. This is crucial for high-stakes applications like medical diagnosis, fraud detection, and cybersecurity, where the cost of a false negative can be catastrophic.integration of generative models (GANs, diffusion models, LLMs) for synthetic data generation is particularly transformative, allowing models to learn from richer, more diverse minority class examples without privacy concerns. Furthermore, the development of sophisticated architectural components and learning paradigms (like attention mechanisms, curriculum learning, and causality-aware designs) ensures that these models are not just accurate but also robust and interpretable.path forward involves continued exploration of hybrid approaches, combining the strengths of various techniques. As models become more complex, the need for robust evaluation frameworks (like those proposed for neonatal seizure detection or link prediction) and explainable AI (XAI) becomes paramount, fostering trust and enabling practical deployment. While challenges like computational efficiency for methods like KANs remain, the progress indicates a future where AI systems can perform reliably even on the most imbalanced and critical real-world datasets, bringing us closer to truly intelligent and equitable AI.