Class Imbalance: Navigating the Uneven Terrain of Modern AI/ML
Latest 11 papers on class imbalance: Nov. 22, 2026
Class imbalance remains a pervasive and challenging issue across diverse AI/ML applications, from medical diagnostics to urban planning and natural language processing. It’s a fundamental problem where certain classes in a dataset are significantly underrepresented compared to others, leading to models that often neglect minority classes or generalize poorly. Recent research, however, is pushing the boundaries, developing innovative strategies to tackle this uneven playing field. This post dives into several recent breakthroughs, revealing how researchers are building more robust, equitable, and efficient AI systems.
The Big Idea(s) & Core Innovations
At the heart of addressing class imbalance is the quest for models that don’t just achieve high overall accuracy but also perform reliably on rare, yet often critical, classes. One significant theme emerging from recent work is the adaptation of existing powerful techniques to explicitly account for imbalance. For instance, in medical image classification, the paper “DAL: Dynamic Angular Loss for Imbalanced Medical Image Classification” by Salman Mohammad, Furkan Kasım, and Manuel Günther from the University of Zurich, introduces a Dynamic Angular Loss (DAL). Their key insight is that standard angular-margin softmax assumptions (like AdaCos) fail when applied to severely imbalanced medical datasets with few classes. They derive data-driven bounds for the softmax scale and angular margin, combining a batch-adaptive angular margin with dynamically weighted cross-entropy. This joint adaptation of representation geometry and optimization weighting significantly boosts minority-class recall and balanced accuracy across challenging medical benchmarks.
Beyond specialized loss functions, the challenge of data streams with delayed labels and concept drift introduces another layer of complexity for imbalanced data. The research by Ahmed Awad and Anirudh Sengar of The British University in Dubai in “When Does Oversampling Help? A Decision Framework for Imbalanced Data Streams Under Label Delay and Concept Drift” systematically investigates the effectiveness of oversampling versus continuous re-evaluation. Their crucial finding: continuous re-evaluation is often the more potent mechanism for recovering performance lost to delayed updates in drift-active streams. Oversampling, while sometimes marginally beneficial, can reinforce outdated concepts if not carefully managed, especially when label delays are long.
Another critical angle is data curation, particularly for multimodal learning. Faidra Patsatzi et al. from Stanford University, in “Reference-Guided Data Curation for 3D Medical Vision-Language Pretraining”, present an ingenious offline, training-free pipeline. Their approach curates medical image-report pairs by retrieving samples aligned with a small, high-quality reference set in both visual and textual embedding spaces, fusing rankings with Reciprocal Rank Fusion. This “quality over quantity” insight allows models to match or exceed full-data performance with significantly less data, enriching rare findings and reducing bias between disease prevalence and downstream performance.
Furthermore, the robustness of models under noisy labels in long-tailed distributions is tackled by Tong Wei et al. from Southeast University and Nanjing University in “Robust Long-Tailed Learning under Label Noise”. They observe that the common “small-loss” criterion for noise detection fails under class imbalance because clean tail-class examples can have high losses due to models’ bias towards head classes. They propose a class-wise small-distance criterion using prototype distances in embedding space and soft pseudo-labeling, leading to improved robustness against both class imbalance and label noise.
In the realm of domain generalization for medical imaging, Zobia Batool et al. from Sabanci University and Boston University, in “Morphology-driven single-domain generalization for cross-cohort Alzheimer’s disease detection using 3D MRIs”, introduce a framework using learnable, class-conditioned morphological augmentations combined with supervised contrastive learning. This method, trained on a single source, generalizes impressively to unseen cohorts for Alzheimer’s detection, highlighting the power of capturing class-specific anatomical variations.
Finally, the evaluation of imbalanced systems itself is a critical area. Jesús S. Aguilar-Ruiz from Pablo de Olavide University, in “The Matthews Correlation Coefficient: Fifty years”, provides a timely retrospective on the Matthews Correlation Coefficient (MCC). This work clarifies MCC’s statistical properties, proving it’s equivalent to Pearson correlation but also demonstrating its sensitivity to class prevalence. This underscores the need for careful metric selection and interpretation when dealing with imbalanced datasets, emphasizing that a single scalar metric might obscure crucial information.
Under the Hood: Models, Datasets, & Benchmarks
These advancements are built upon and contribute to a rich ecosystem of models, datasets, and benchmarks:
- Dynamic Angular Loss (DAL): Utilizes ResNet-18, DenseNet121, and ResNet50 backbones on highly imbalanced medical imaging benchmarks like ISIC 2018/2019, APTOS 2019, and Hyper-Kvasir. The official DAL source code is available at https://github.com/msalman95/DAL.
- Streaming Oversampling Decision Framework: Compares 15 classifiers, including streaming oversampling variants (e.g., SMOTE, VFC-SMOTE, C-SMOTE) and bare streaming learners, on six synthetic and six real-world data streams (e.g., Electricity, KDDCup99, Credit Card Fraud). Code and results are at https://github.com/AhmedAwad/streaming-oversampling-under-delay-drift/.
- Reference-Guided Data Curation: Employs DreamSim embeddings for images and Clinical Longformer for radiology reports on 3D abdominal CT (Merlin Abdominal CT dataset) and 2D chest X-rays (CheXpert Plus, MIMIC-CXR). The code for curation, training, and evaluation is open-sourced at https://github.com/StanfordMIMI/data-curation-3d-medical-vlm.
- Robust Long-Tailed Learning: Evaluated on CIFAR-10/100, WebVision, mini WebVision, and ImageNet-LT, using prototypes and Gaussian mixture models for noise detection.
- Morphology-driven Generalization: Leverages 3D U-Net encoder, trained on NACC, and tested directly on ADNI, AIBL, and OASIS-3 cohorts for Alzheimer’s detection. The code is available at https://github.com/zobia111/MorphoSDG.
- TabBioMed: A large-scale benchmark introduced by Pau Mateo-Bernadó et al. from Stanford University and Universitat Politècnica de Catalunya, curating 94 biomedical tabular datasets across eight domains. It benchmarks 23 models, including classical baselines, gradient-boosted trees, neural tabular architectures, and tabular foundation models. The datasets, splits, and code are on Hugging Face (https://huggingface.co/spaces/TabBioMed) and GitHub (https://github.com/AI-sandbox/TabBioMed).
- Video Capsule Endoscopy Evaluation: Smriti Regmi and Debesh Jha from the University of South Dakota introduce DMSNet (an EfficientNetV2-S-based architecture) and benchmark it against 18 CNN/transformer models on Kvasir-Capsule and the independent CV2024 test set. The paper is “Rethinking Evaluation for Video Capsule Endoscopy: Validation Protocol, Source-Video Diversity, and Cross-Dataset Shift”.
- Turkish Multimodal Video Fact-Checking: Elnur Adl Zarabi et al. from Texas A&M University-Corpus Christi introduce TeViD, a 2,097-example dataset for Turkish multimodal video fact-checking. They compare zero-shot agentic pipelines with LoRA fine-tuned Qwen2.5-7B models. The code and resources are at https://github.com/PASHAADC/Danish-LLM-HateSpeech and https://github.com/MeiTaylor/TRUE-3MFact.
- HFX3D: Abdiaziz Muse et al. from Saint Mary’s University introduce “From Geometry to Function: A Multi-Task LiDAR Benchmark Dataset for Semantic and Functional Parsing of Urban Facades”. This benchmark offers approximately 492 million RGB-colored LiDAR points from urban buildings, annotated for semantic, instance, and functional attributes. The dataset and benchmark code are available at https://jijup.github.io/hfx3d-benchmark/ and https://github.com/jijup/hfx3d-benchmark.
Impact & The Road Ahead
These diverse advancements demonstrate a concerted effort to move beyond simplistic accuracy metrics and build AI systems that are genuinely robust and fair in the face of real-world data challenges. The implications are far-reaching: from more reliable medical diagnoses for rare conditions, to more accurate fraud detection in streaming data, and more effective content moderation in low-resource languages.
The push for rigorous evaluation protocols, such as source-video-disjoint validation in “Rethinking Evaluation for Video Capsule Endoscopy: Validation Protocol, Source-Video Diversity, and Cross-Dataset Shift” or the multifaceted TabBioMed benchmark for biomedical tabular data, is crucial. It highlights that strong in-domain performance doesn’t guarantee real-world utility and that context-aware evaluation is paramount.
Looking ahead, the synergy between data curation, robust loss functions, and domain generalization techniques will continue to be vital. The success of zero-shot LLMs with careful prompting and lightweight fine-tuning for low-resource tasks, as shown in “On the Effectiveness and Limitations of Open-Weight LLMs for Low-Resource Offensive Language Detection in Danish”, signals a promising direction for deploying powerful models in challenging linguistic contexts despite imbalance. The ongoing journey to tame class imbalance will lead to more trustworthy, equitable, and ultimately, more intelligent AI systems for everyone.
Share this content:
Discover more from SciPapermill
Subscribe to get the latest posts sent to your email.
Post Comment