Class Imbalance: Bridging the Gap Between Scarce Data and Robust AI

Latest 22 papers on class imbalance: Jul. 18, 2026

Class imbalance remains a pervasive and critical challenge across diverse AI/ML applications, from medical diagnostics to fraud detection and cybersecurity. When one class significantly outnumbers others, models often struggle to accurately identify the minority class, leading to biased predictions, missed anomalies, and compromised real-world utility. Fortunately, recent research offers a fascinating array of breakthroughs, leveraging everything from quantum computing to advanced prompt engineering, to tackle this fundamental problem head-on.

The Big Idea(s) & Core Innovations

Many recent innovations center around smart data augmentation, adaptive model training, and rethinking evaluation metrics. For instance, the NodeImport framework by Nan Chen et al. from Johns Hopkins University and National University of Singapore introduces a novel approach for class-imbalanced node classification in graphs. Their core insight is to identify “important” nodes—those that improve model performance in an unbiased setting—using a balanced meta-set and a tractable importance formula. This allows for filtering valuable labeled, unlabeled, and synthetic nodes, showcasing that carefully curated data, rather than just more data, is key.

Similarly, for tabular data, the paper “Quantum-Enhanced Synthetic Data Generation Using Quantum Circuit Born Machines for Imbalanced Tabular Learning” by Tanapol Nuatho et al. from King Mongkut’s University of Technology Thonburi explores a hybrid quantum-classical framework. They find that Quantum Circuit Born Machines (QCBMs) can generate synthetic data with superior structural fidelity (lower MMD) compared to classical SMOTE variants, leading to significant F1-score and minority-class recall improvements. This suggests quantum entanglement might capture complex cross-feature correlations better than traditional methods.

Addressing the generation of useful synthetic data, Yanxuan Yu et al. from Columbia University and UCLA propose RUBRIC: Realism–Utility Balanced Ranking for Imbalanced Classification. This generator-agnostic post-filtering framework selects synthetic samples based on a balance of realism (quantified by a discriminator) and utility (proximity to decision boundaries). Their work highlights that filtering synthetic samples based on these criteria can tighten generalization bounds and improve F1/recall, moving beyond merely generating more data to generating better data.

Beyond data generation, model adaptation is crucial. In “Parameter-efficient Prompt Tuning of Vision Foundation Model With Adaptive Focal Loss for Interpretable MCI Screening”, Javad Khoramdel et al. from K. N. Toosi University of Technology demonstrate that frozen vision foundation models with lightweight, task-specific prompt tuning can outperform computationally expensive architectures for medical imaging tasks like MCI screening. Their adaptive focal loss and MoCA-probability modulator encode diagnostic uncertainty directly into loss gradients, proving more effective than soft labels alone. This echoes the finding in “Mitigating The Effect of Class Imbalance in Data with Hierarchical and Dependable Structure” by Bipin Chhetri et al. from Texas Tech University, where synthetic oversampling degrades deep learning performance on hierarchical data like CWE vulnerability classification, advocating instead for Hierarchy-Aware RoBERTa that preserves taxonomic consistency.

For vision-language models, Francis Fernandez et al. from San Diego State University present C-GAP, an annotation-free, training-free iterative caption refinement pipeline. Their insight is that prompt sensitivity, not detector capacity, often bottlenecks minority-class detection. By refining text prompts based on detector feedback, they achieve substantial relative improvements in minority-class AP@0.5 without updating any model weights, a critical advancement for safety-critical perception.

In network intrusion detection, Abu Fuad Ahmad and Istiaque Ahmed from New Mexico State University and Osaka Metropolitan University introduce BARS: Benign-Anchored Ranking and Selection. They address a structural failure mode in feature selection methods where global anchors drift towards the dominant class under imbalance. By using the benign-class mean as an anchor, BARS achieves 15-23% FPR reduction on attack-majority datasets, proving that re-anchoring can yield significant gains.

Finally, for computational efficiency, Jean-François Bonbhel from NeuroSoft IA proposes K-ABENA, a selective backpropagation framework that saves 28-54% of gradient computation while maintaining unbiased gradient estimation. A crucial finding is that uncompensated selective backpropagation methods structurally fail under extreme class imbalance, highlighting the necessity of K-ABENA’s inverse-probability weighting with a defensive mixture design.

Under the Hood: Models, Datasets, & Benchmarks

The papers introduce or heavily utilize a range of models, datasets, and benchmarks that are crucial for advancing research in class imbalance:

• DINOv2-Small & MoCA-aware Focal Loss: Utilized by Khoramdel et al. for MCI screening on neuropsychological drawing tests, demonstrating parameter-efficient adaptation with intrinsic interpretability. Code: https://github.com/JVD9kh96/mci-detection.
• XCT-SAM & Conv-LoRA: Developed by Md Mahedi Hasan et al. from West Virginia University for industrial XCT defect segmentation, leveraging a two-stage domain adaptation strategy with only 0.647% of SAM’s parameters. Code: https://github.com/Mahedi-61/XCT-SAM.git.
• QFireNet & Variational Quantum Circuits: Introduced by Jaiman Munshi et al. from University of Maryland as a hybrid quantum-classical U-Net for wildfire segmentation from Sentinel-2 imagery, outperforming classical baselines.
• NodeImport & Balanced Meta-set Selection: Nan Chen et al. provide a framework for imbalanced node classification on Graph Neural Networks, with code available at https://github.com/NanChanNN/NodeImport.
• BARS & NIDS Datasets (UNSW-NB15, CICDDoS2019): Abu Fuad Ahmad and Istiaque Ahmed refined feature selection for Network Intrusion Detection Systems using benign-anchoring, showing effectiveness on attack-majority datasets. Code to be released.
• Hierarchy-Aware RoBERTa & CWE Taxonomy: Bipin Chhetri et al. developed this for cybersecurity vulnerability classification, demonstrating the pitfalls of oversampling on hierarchical data.
• DiffEEG & Denoising Diffusion Models: Abdulkader Helwan et al. from Lebanese American University and TÉLUQ University proposed a self-supervised foundation model for EEG seizure detection, pre-trained on the Temple University Hospital Seizure Corpus, and fine-tuned with reinforcement learning.
• GFD-GC & Grouped Attribute Completion: Junpeng Wu and Ye Yuan from Southwest University addressed graph fraud detection on real-world datasets like Amazon and Yelp, combining attribute completion with confidence-aware contrastive learning.
• CatBoost & Telco Customer Churn Dataset: Nada Ali et al. from University of Khartoum presented an integrated ML framework for telecom churn prediction and customer segmentation, with code available upon request.
• SMETA-ZSL & Cyber Threat Intelligence: Ivan Alejandro Montoya Sanchez et al. from The University of Texas at El Paso introduced this framework for zero-shot threat classification in cybersecurity, leveraging LLMs and meta-learning. Code: https://github.com/Security-And-Intelligence-Lab-UTEP/SMETA-ZSL.
• Temperature Scaling & Mendeley LBC Dataset: Nisreen Albzour and Sarah S. Lam from Binghamton University highlighted the importance of post-hoc calibration over ensemble size for reliability in cervical cytology classification.
• SHAL & Whole Slide Images (TCGA colorectal cancer): Mahsa Vali et al. from University of Cologne developed a patient-level active learning framework for histopathology, reducing annotation burden by 64%.
• RUBRIC & Fraud Detection Benchmarks: Yanxuan Yu et al. evaluated their synthetic sample selection method on datasets like Credit Card Fraud Detection (ULB) and IEEE-CIS Fraud Detection. Code: https://github.com/zorinayu/RUBRIC.
• Semantic Pareto-DQN & E-Commerce Fraud: Cláudio Lúcio do Val Lopes and Lucca Machado da Silva from A3Data presented a multi-objective reinforcement learning framework that breaks the “fraud collapse” trap in financial anomaly detection.
• JEFFNet & PVF-10 Dataset: Seyyedhamid Azimidokht et al. from University of Oulu proposed a multibranch architecture for solar PV panel fault classification. Code: https://github.com/Azimi2kht/JEFFNet.
• QCBM & Iris, Telco Customer Churn Datasets: Tanapol Nuatho et al. extensively benchmarked their quantum-enhanced synthetic data generation against classical methods. Relies on Qiskit and scikit-learn.
• C-GAP & COCO, Cityscapes, Chula Vista datasets: Francis Fernandez et al. utilized established vision-language models like Grounding DINO and YOLO-World for their prompt refinement work.
• DeepPySR & Biomedical/Social Science Data: Fuling Chen et al. from University of Western Australia extended symbolic regression to handle high-dimensional, imbalanced data, producing interpretable formulas. Code: https://github.com/ICRAR/deeppysr.jl.
• DualAlign & MM-JDM Dataset: Kanglei Zhou et al. from Tsinghua University introduced a new multi-modal movement-quality assessment dataset for Action Quality Assessment (AQA), along with their two-stage fusion framework.
• VendorBench-100: Sharayu N. Deshmukh et al. from Universidade da Beira Interior created a unified benchmark for deepfake detection, revealing critical divergences between ranking ability and operating-point quality. Code: https://github.com/sharayu-20/vendorbench-100.
• K-ABENA & CIFAR-10, ImageNet (implicit): Jean-François Bonbhel validated his selective backpropagation framework’s efficiency and unbiasedness. Code: https://github.com/Bonbhel/kabena-ml.

Impact & The Road Ahead

These advancements have profound implications. The ability to robustly handle class imbalance, often through parameter-efficient adaptation, self-supervised learning, or novel optimization techniques, unlocks new possibilities for deploying AI in critical, data-scarce domains. Imagine more accurate early disease detection with minimal labeled data, highly reliable fraud prevention, or enhanced cybersecurity systems capable of identifying emerging threats without prior examples. The trend towards integrating semantic understanding (via LLMs) with structured data and visual features further promises models that are not only more accurate but also more interpretable and adaptable. Future work will likely focus on pushing the boundaries of quantum-classical hybrids, developing more sophisticated mechanisms for balancing utility and realism in synthetic data, and creating universally robust adaptive frameworks that can operate effectively even with zero-shot generalization. The continuous innovation in addressing class imbalance is not just refining existing AI; it’s fundamentally expanding where and how AI can make a meaningful impact.

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.