Class Imbalance No More: Recent Breakthroughs in AI/ML Tackle the Skewed Data Challenge

Latest 50 papers on class imbalance: Dec. 21, 2025

Class imbalance — a pervasive problem where some categories have significantly fewer samples than others — continues to be a formidable challenge in AI and Machine Learning. From rare disease detection to spotting obscure cyberattacks, this skewed data distribution can severely hobble model performance, leading to biased predictions and overlooked critical events. But fear not, for recent research is bringing forth innovative solutions that are revolutionizing how we handle this thorny issue, paving the way for more robust, fair, and accurate AI systems.

The Big Idea(s) & Core Innovations

At the heart of these advancements is a multi-pronged attack on class imbalance, leveraging everything from novel loss functions and data augmentation to sophisticated ensemble methods and generative models. A key insight emerging from several papers is that simply re-weighting classes or oversampling isn’t enough; more nuanced, context-aware strategies are required. For instance, the paper “The Multiclass Score-Oriented Loss (MultiSOL) on the Simplex” by Francesco Marchetti, Edoardo Legnaro, and Sabrina Guastavino from the University of Padova and Genova introduces MultiSOL, a novel family of loss functions extending score-oriented losses to multiclass settings. This allows for direct optimization of target performance metrics, offering a more robust approach in imbalanced scenarios.

In the realm of functional data, “Functional Random Forest with Adaptive Cost-Sensitive Splitting for Imbalanced Functional Data Classification” by Fahad Mostafa and Hafiz Khan, affiliated with Arizona State University and Texas Tech, proposes FRF-ACS. This innovative ensemble method integrates basis expansions, adaptive cost-sensitive splitting, and hybrid resampling to significantly improve minority-class detection while preserving functional geometry – crucial for applications like ECG analysis.

Cybersecurity, often plagued by rare attack instances, sees significant advancements. “PHANTOM: Progressive High-fidelity Adversarial Network for Threat Object Modeling” by Jamal Al-Karaki, Muhammad Al-Zafar Khan, and Rand Derar Mohammad Al Athamneh from Zayed University and The Hashemite University, introduces an adversarial variational framework for generating synthetic cyberattack datasets. PHANTOM’s progressive training and dual-path learning create realistic data, enhancing intrusion detection despite inherent scarcity. Similarly, the “Hybrid Ensemble Method for Detecting Cyber-Attacks in Water Distribution Systems Using the BATADAL Dataset” combines Random Forest, XGBoost, and LSTM, demonstrating that a hybrid stacked ensemble framework can significantly outperform individual models by effectively handling class imbalance and temporal dependencies, with interpretability boosted by SHAP analysis.

In computer vision, especially with visual language models, representation entanglement in long-tailed distributions is a critical issue. “CORAL: Disentangling Latent Representations in Long-Tailed Diffusion” from researchers at Arizona State University introduces CORAL, a contrastive latent alignment method. By using supervised contrastive loss, CORAL disentangles latent representations, dramatically improving the diversity and visual fidelity of samples from underrepresented classes in diffusion models. This concept extends to medical imaging with “Pretraining Transformer-Based Models on Diffusion-Generated Synthetic Graphs for Alzheimer’s Disease Prediction” by Abolfazl Moslemi and Hossein Peyvandi, where diffusion-based synthetic data generation mitigates label imbalance and data scarcity for improved early Alzheimer’s detection using Graph Transformers.

For LLMs, proper evaluation metrics are vital when dealing with judges on imbalanced data. “Balanced Accuracy: The Right Metric for Evaluating LLM Judges – Explained through Youden’s J statistic” by Stephane Collot et al. from Meta Superintelligence Labs argues for Balanced Accuracy, showing its robustness against prevalence-dependent metrics like F1 or Precision, particularly in skewed settings.

Under the Hood: Models, Datasets, & Benchmarks

These innovations are often underpinned by specialized models, datasets, and benchmarks that push the boundaries of what’s possible in real-world, imbalanced scenarios:

• PHANTOM’s Adversarial Variational Framework: Utilizes progressive training and dual-path learning (VAE stability + GAN fidelity) to generate high-fidelity synthetic cyberattack data, achieving 98% weighted accuracy in intrusion detection.
• BATADAL Dataset: Heavily utilized by the “Hybrid Ensemble Method for Detecting Cyber-Attacks in Water Distribution Systems Using the BATADAL Dataset” for evaluating hybrid stacked ensembles (Random Forest, XGBoost, LSTM) against real-world industrial control system threats. Code for the framework is noted as available.
• CORAL’s Supervised Contrastive Loss: Applied within U-Net architectures, specifically targeting the bottleneck layer, to disentangle latent representations for long-tailed diffusion models. The authors provide a GitHub repository: https://github.com/SankarLab/coral-lt-diffusion.
• RCDGen for Remote Sensing: “Referring Change Detection in Remote Sensing Imagery” introduces RCDGen, a synthetic data generation pipeline based on diffusion models. This addresses class imbalance and data scarcity in remote sensing by generating realistic post-change images without needing semantic segmentation masks. Code is available at https://github.com/huggingface/.
• LNMBench: Introduced in “Benchmarking Real-World Medical Image Classification with Noisy Labels” by Yuan Maa et al., this is a comprehensive benchmark for evaluating robustness under label noise in medical image classification across multiple datasets and modalities. Code is publicly available at https://github.com/myyy777/LNMBench.
• DermETAS-SNA LLM: Introduced in “DermETAS-SNA LLM: A Dermatology Focused Evolutionary Transformer Architecture Search with StackNet Augmented LLM Assistant” by Santosh et al., this system uses an Evolutionary Transformer Architecture Search (ETAS) and StackNet augmented LLMs. It achieves significant F1-score increases (16.06%) over baselines like SkinGPT-4 for dermatological diagnosis, especially for rare conditions.
• CICLe Framework: From “Efficient Text Classification with Conformal In-Context Learning” by Ippokratis Pantelidis et al. at Stockholm University, this resource-efficient framework combines conformal prediction with lightweight base classifiers, performing well in highly imbalanced text classification scenarios. Code is available at https://github.com/ippokratis-pantelidis/CICLe.
• SD-CGAN for IoT Security: “SD-CGAN: Conditional Sinkhorn Divergence GAN for DDoS Anomaly Detection in IoT Networks” leverages conditional Sinkhorn divergence with GANs to enhance DDoS attack detection, improving robustness against subtle attack patterns.
• ASD for Federated Learning: “Adaptive Self-Distillation for Minimizing Client Drift in Heterogeneous Federated Learning” introduces Adaptive Self-Distillation (ASD), a computationally efficient regularization method. It mitigates client drift by adaptively weighting regularization loss based on global model predictions and local label distributions, with code at https://github.com/vcl-iisc/fed-adaptive-self-distillation.
• Hybrid SDG Framework for Zero-Shot Inspection: The paper “Hybrid Synthetic Data Generation with Domain Randomization Enables Zero-Shot Vision-Based Part Inspection Under Extreme Class Imbalance” proposes a groundbreaking hybrid synthetic data generation framework for industrial part inspection. It achieves 90–91% balanced accuracy using only synthetic data under severe class imbalance (11:1 pass/fail), eliminating the need for manual annotation in manufacturing quality control.

Impact & The Road Ahead

The implications of these advancements are profound. By effectively mitigating class imbalance, AI models can become more trustworthy in critical domains like healthcare, where missing rare but severe conditions can have dire consequences. In cybersecurity, these methods enable better detection of sophisticated, infrequent attacks, bolstering our digital defenses. For industrial automation, zero-shot learning with synthetic data promises faster deployment and significant cost savings in quality control.

Looking ahead, the synergy between generative models, specialized loss functions, and interpretable AI will continue to deepen. We can expect more sophisticated adaptive strategies that not only handle imbalance but also inherently understand the contextual significance of minority classes. The emphasis on robust benchmarking and clear evaluation metrics like Balanced Accuracy will further guide research towards truly impactful solutions. As AI continues to integrate into sensitive, real-world applications, addressing class imbalance isn’t just a technical detail—it’s a critical step towards building AI that is reliable, fair, and truly intelligent.