Review Article
A comprehensive survey on imbalanced data learning
This survey provides a comprehensive and up-to-date overview of imbalanced data learning, organizing existing research into four categories: data re-balancing, feature representation, training strategy, and ensemble learning. It delves into three critical real-world data formats (images, text, graphs), highlighting their characteristics and unique challenges. The paper emphasizes imbalance learning as both a foundational component and a beneficiary of modern machine learning advancements, especially with the rise of generative AI and large-scale pretraining models. It aims to bridge classical techniques with modern paradigms, guiding researchers and practitioners towards scalable, modality-aware imbalance learning methods for robust, fair, and adaptive intelligent systems.
Executive Impact & Key Takeaways
Understand the critical implications of imbalanced data and the strategic advantages of advanced AI solutions.
0
Datasets Evaluated
0
Methods Compared
0
Performance Improvement (F1-score Avg)
Key Insights from the Survey:
- Imbalanced data hinders ML performance by biasing decision-making, prevalent in fraud detection (0.17% fraud), healthcare (rare diseases <1%), and manufacturing (defects rare by design).
- Solutions are categorized into data re-balancing, feature representation, training strategy, and ensemble learning, enabling targeted research and development.
- The survey highlights the pervasive nature of imbalance across diverse data formats like images, text, and graphs, each with unique challenges and required adaptations.
- Open-source libraries like imbalanced-learn and IGL-Bench are crucial resources for researchers and practitioners, offering structured analysis and benchmarks.
- Future directions include addressing imbalance in open-world settings, multi-label problems, imbalanced regression, multi-modality contexts, and leveraging Large Language Models (LLMs) for data synthesis and adaptation.
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Data Re-balancing Overview
Data re-balancing methods modify input data distribution to mitigate class imbalance during data preparation. This includes over-sampling minority classes, under-sampling majority classes, and hybrid approaches to achieve a more balanced dataset for training.
100+
SMOTE variants developed, most influential method in data generation.
SMOTE (Synthetic Minority Oversampling Technique) selects k nearest minority neighbors and performs linear interpolation to create new synthetic minority samples. It effectively mitigates overfitting and enhances classifier learning. Over 100 variants exist, addressing issues like over-representation and density distribution.
Data Re-balancing Flow
Data generation (e.g., SMOTE, GANs)
→
Adaptive down-sampling (e.g., Tomek Link, CNN)
→
Hybrid sampling (combining over/under-sampling)
→
Re-labeling (e.g., Self-training, Active learning)
| Method Category | Key Advantages | Limitations |
|---|---|---|
| Linear Generative (e.g., SMOTE) |
|
|
| Deep Generative (e.g., GANs, VAEs, Diffusion Models) |
|
|
Feature Representation Overview
These methods aim to learn discriminative embeddings that better capture minority class characteristics, improving class separability in the latent space. Techniques include cost-sensitive learning, metric learning, supervised contrastive learning, prototype learning, transfer learning, and meta-learning.
Focal Loss
Assigns higher costs to challenging minority samples.
Focal Loss addresses imbalance by assigning varying weights to each training sample, giving higher costs to difficult-to-classify minority samples and lower costs to easy-to-classify ones. This helps models focus on critical examples.
Feature Representation Flow
Cost-sensitive learning (e.g., Focal Loss)
→
Metric learning (e.g., Triplet Loss)
→
Supervised contrastive learning (e.g., SupCon)
→
Prototype learning (e.g., ProAug)
→
Transfer learning (e.g., FTL)
→
Meta learning (e.g., Meta-Weight-Net)
Case Study: Metric Learning for Fault Diagnosis
Context: Imbalanced time-series data is common in fault diagnosis, where healthy operational states far outnumber fault states. Traditional models often misclassify rare faults due to biased feature spaces.
Approach: A quadruplet deep metric learning model (LSTM-QDM) was applied. It leverages minority samples to expand data triplets and uses a quadruplet loss to balance relationships among anchor, positive, negative, and minority samples. This approach also jointly optimizes softmax loss.
Result: LSTM-QDM achieved a stronger representation ability for imbalanced time-series data, effectively separating fault classes and improving detection accuracy for rare faults. The combined optimization enhanced both within-class compactness and between-class separation.
Training Strategy Overview
Training strategies adjust the learning process to reduce bias toward the majority class during model training. This includes decoupling training (separating representation learning from classification), fine-tuning pre-trained models, curriculum learning (learning from easy to hard), and posterior recalibration.
Decoupled Training
Separates representation learning and classifier training.
Decoupled training addresses imbalance by separating the model's learning into two stages: representation learning and classification. This allows the model to first learn high-quality, unbiased features and then train the classifier on a balanced subset.
Training Strategy Flow
Decoupling training
→
Fine-tuning pre-trained model
→
Curriculum learning
→
Posterior recalibration
Case Study: Dynamic Curriculum Learning for Image Classification
Context: Long-tailed image datasets, where a few classes have many images and many classes have few, pose a significant challenge. Models tend to perform poorly on rare classes.
Approach: Dynamic Curriculum Learning (DCL) was used. It integrates curriculum learning into a sampling scheduler, dynamically transitioning the training from imbalanced to balanced and from easy to hard examples. It also incorporates metric learning to create a soft feature embedding.
Result: DCL improved generalization by first learning suitable feature representations and then shifting the learning objective from metric learning to classification loss. This dynamic adaptation enhanced model performance on both majority and minority classes, particularly for rare classes.
Ensemble Learning Overview
Ensemble methods integrate multiple models to enhance robustness and generalization, leveraging the diversity of learners to mitigate data imbalance effects during inference. Techniques include bagging-based, boosting-based, cost-sensitive, and knowledge distillation ensembles.
SMOTE-Bagging
Combines SMOTE with Bagging for diverse synthetic samples.
SMOTE-Bagging integrates the SMOTE oversampling technique into the Bagging algorithm. It generates synthetic minority samples through interpolation, ensuring diversity in the training sets for base classifiers and reducing overfitting.
Ensemble Learning Flow
Bagging-based methods (e.g., SMOTE-Bagging)
→
Boosting-based methods (e.g., RAMOBoost)
→
Cost sensitive-based methods (e.g., AdaCost)
→
Knowledge distillation (e.g., Multi-teacher KD)
Case Study: RUSBoost for Medical Diagnosis
Context: In medical diagnosis, correctly identifying rare but critical diseases (minority class) is crucial, but traditional classifiers are biased towards abundant healthy cases (majority class).
Approach: RUSBoost, a hybrid approach combining Random Under-Sampling (RUS) with AdaBoost, was applied. RUS randomly selects samples from the majority class in each boosting iteration, while AdaBoost iteratively re-weights samples, focusing on misclassified ones.
Result: RUSBoost successfully enhanced the detection of minority class instances by reducing the bias towards the majority class. While simple to implement and fast, it effectively improved overall classification performance in imbalanced medical datasets.
Calculate Your Potential AI ROI
Estimate the potential cost savings and efficiency gains your organization could achieve by implementing optimized AI solutions for imbalanced data.
Future Directions & Your AI Roadmap
Pioneering the next generation of AI solutions requires foresight into emerging challenges and advanced methodologies. Our roadmap outlines key areas for strategic development.
Imbalance in Open-World Settings
Models must adapt to emerging, unknown classes in test sets. This requires uncertainty-aware learning frameworks combined with open-set recognition and continual learning, and using generative models to simulate novel class distributions.
Imbalanced Multi-Label Problem
Challenges arise from label overlap and semantic entanglement across labels. Solutions involve refined loss functions, label-specific re-sampling, and cross-modal augmentation to handle diverse label distributions.
Imbalanced Regression Problem
Continuous labels lack clear minority/majority boundaries. Future research should focus on structure-aware solutions that preserve relative distances, ranking-based objectives, and distribution-aware sampling emphasizing rare but critical value regions.
Multi-Modality Imbalance Problem
Modalities differ in data volume, quality, or annotation. Need imbalance-aware fusion strategies and cross-modal augmentation to address representational disparities and prevent over-reliance on dominant modalities.
Large Language Models (LLMs) for Imbalance
LLMs can generate class-conditional, semantically coherent data for minority classes. LLM-based augmentation through prompt design and meta-learning with LLMs are promising for scalable and adaptive solutions.
Ready to Transform Your Data Strategy?
Leverage cutting-edge insights to overcome data imbalance and drive superior AI performance. Book a free consultation with our experts to design a tailored solution for your enterprise.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat