[Paper Review] Class-Imbalanced Semi-Supervised Learning

Minsung Hyun, Jisoo Jeong+1 more|arXiv (Cornell University)|Feb 17, 2020

Imbalanced Data Classification Techniques42 references37 citations

TL;DR

Defines the task CISSL (class-imbalanced semi-supervised learning) and proposes Suppressed Consistency Loss (SCL) to improve SSL when both labeled and unlabeled data are imbalanced, showing MT+SCL often outperforms standard SSL methods especially with severe imbalance and limited labeled data.

ABSTRACT

Semi-Supervised Learning (SSL) has achieved great success in overcoming the difficulties of labeling and making full use of unlabeled data. However, SSL has a limited assumption that the numbers of samples in different classes are balanced, and many SSL algorithms show lower performance for the datasets with the imbalanced class distribution. In this paper, we introduce a task of class-imbalanced semi-supervised learning (CISSL), which refers to semi-supervised learning with class-imbalanced data. In doing so, we consider class imbalance in both labeled and unlabeled sets. First, we analyze existing SSL methods in imbalanced environments and examine how the class imbalance affects SSL methods. Then we propose Suppressed Consistency Loss (SCL), a regularization method robust to class imbalance. Our method shows better performance than the conventional methods in the CISSL environment. In particular, the more severe the class imbalance and the smaller the size of the labeled data, the better our method performs.

Motivation & Objective

• Motivate and formalize the class-imbalanced semi-supervised learning (CISSL) setting where imbalance exists in both labeled and unlabeled data.
• Analyze how existing SSL methods behave under CISSL and identify failure modes caused by imbalance.
• Introduce Suppressed Consistency Loss (SCL) to regularize SSL in CISSL and demonstrate robustness across imbalance scenarios.
• Extend CISSL to object detection and show gains over baseline SSL approaches.

Proposed method

• Define CISSL and standard experimental settings reflecting balanced and imbalanced distributions in labeled and unlabeled data.
• Analyze existing SSL methods (Pi model, Mean Teacher) under CISSL using toy datasets to understand boundary behavior.
• Propose Suppressed Consistency Loss (SCL) that weights the SSL consistency loss inversely to class frequency via g(Nc) with an exponential form.
• Demonstrate empirical improvements of MT+SCL over baselines on CIFAR-10 and SVHN under various unlabeled imbalance types.
• Extend SCL to an object detection task (SSD300/CSD framework) and show performance gains.

Experimental results

Research questions

• RQ1 How do standard SSL methods perform when both labeled and unlabeled data are class-imbalanced (CISSL)?
• RQ2 Why is Mean Teacher generally more robust than Pi model under CISSL, and how does target updating influence this behavior?
• RQ3 Does introducing Suppressed Consistency Loss mitigate degradation in minor classes and improve overall CISSL performance?
• RQ4 Do the CISSL gains of SCL extend to other tasks like object detection?

Key findings

• SSL methods based on consistency regularization degrade on minor classes under CISSL; Mean Teacher is more robust than the Pi model due to its conservative EMA target.
• Suppressed Consistency Loss (SCL) mitigates boundary distortion in minor classes by inversely weighting the consistency loss with class frequency, improving performance across CIFAR-10 and SVHN under imbalanced unlabeled data.
• MT+SCL consistently achieves the best or near-best results across multiple CISSL configurations, especially as imbalance and unlabeled imbalance increase.
• In ablations, standard class-imbalanced learning losses (IN, Focal, CB) do not universally outperform CE in CISSL, highlighting the need for CISSL-specific regularization.
• Applying SCL to object detection (CSD framework) yields additional gains over the baseline CSD method.