Survey on deep learning with class imbalance

Analysis

Why This Paper Matters

Class imbalance is a pervasive challenge in real-world AI applications, from fraud detection to medical diagnosis, where minority classes often carry the highest stakes. While traditional machine learning has extensively studied imbalance, deep learning's rise has left a critical gap. This 2019 survey by Johnson and Khoshgoftaar, with nearly 3000 citations, systematically maps that gap, making it an essential reference for any practitioner deploying neural networks on skewed data. The paper's timing is key: it captures the state of the field just as deep learning was becoming ubiquitous, providing a baseline for subsequent advances.

The survey's value lies in its structured analysis of data complexity, architectures, and performance interpretation. It confirms that many classic remedies—data sampling, cost-sensitive learning—remain effective in deep learning contexts, but also points to emerging techniques that leverage neural networks' feature learning abilities. For practitioners, this means not abandoning proven methods while staying alert to newer, more powerful approaches.

Technical Contributions

• Comprehensive taxonomy: Categorizes methods into data-level (oversampling, undersampling), algorithm-level (cost-sensitive losses), and hybrid approaches, with specific attention to deep learning adaptations.
• Focus on CNNs: Identifies that the vast majority of existing work uses convolutional neural networks for computer vision tasks, highlighting a narrow scope.
• Evaluation of traditional methods: Demonstrates that techniques like SMOTE and cost-sensitive learning transfer to deep learning, but with caveats about computational cost and scalability.
• Advanced methods: Reviews approaches that modify network architectures (e.g., focal loss, class-balanced sampling) to directly address imbalance during training.
• Gap analysis: Explicitly notes the lack of research on big data effects, generalization to non-vision domains, and rigorous performance metrics beyond accuracy.

Results

The survey does not present new experimental results but synthesizes findings from dozens of studies. Key takeaways include: (1) traditional data sampling methods remain effective but can introduce noise; (2) cost-sensitive learning is straightforward to implement and often improves minority class recall; (3) advanced methods like focal loss show strong results on highly imbalanced vision benchmarks; (4) most studies report accuracy or F1-score, with few using metrics like AUC or precision-recall curves that are more appropriate for imbalance. The paper emphasizes that research is "very limited" and that big data considerations are "rarely considered."

Significance

This survey has become a cornerstone reference for anyone working on class imbalance in deep learning. Its clear articulation of gaps—especially the need for work beyond computer vision and with large-scale data—has directly influenced subsequent research directions. For practitioners, it provides a practical roadmap: start with simple sampling or cost-sensitive modifications, then explore advanced feature-learning methods if needed. The paper's high citation count reflects its utility as both a literature review and a call to action, cementing its role in shaping the field's evolution toward more robust, fair, and effective deep learning systems.