Active learning promises to minimize annotation costs by selecting only the most informative samples from an unlabeled pool. In an ideal scenario, maximizing predictive uncertainty (such as entropy) guides the model directly to the decision boundary. However, real-world datasets are rarely balanced. When a dataset is dominated by a single majority class, vanilla active learning strategies suffer from a critical failure mode: they get trapped optimizing the nuances of the majority class while completely missing rare, safety-critical minority categories.
To overcome this, a prominent research direction focuses on uncertainty-driven class balancing. By tightly coupling a model's predictive uncertainty with explicit class-frequency metrics, these methods reshape the data acquisition loop to flatten the long tail of imbalanced datasets.
The Dual Nature of Uncertainty Under Imbalance
To understand why class-balanced active learning works, we must first look at how class imbalance distorts standard uncertainty metrics.
When a model is trained on an imbalanced dataset, it naturally exhibits high epistemic uncertainty for minority classes simply because it has seen fewer examples of them. As Chaitanya Shivade et al. observed in Addressing Limited Data for Textual Entailment Across Domains, standard uncertainty sampling can sometimes act as an implicit balancing mechanism because the model disproportionately selects rare positive instances it is unsure about.
However, this implicit effect is fragile. If the majority class contains significant noise, out-of-distribution artifacts, or complex sub-distributions, its aleatoric uncertainty will dwarf the epistemic uncertainty of the minority class. In these cases, a purely uncertainty-driven approach will over-sample the majority class due to its sheer volume in the unlabeled pool, completely neglecting rare categories.
Explicit Class-Aware Weighting
The most direct way to prevent majority-class dominance is to inject class-frequency awareness directly into the acquisition function. The core intuition is to scale a sample's uncertainty score inversely by the prevalence of its predicted class in the currently labeled pool:
$$\text{Acquisition Score}(x) \propto \text{Uncertainty}(x) \times \frac{1}{\text{Frequency}(\hat{y})}$$
This soft weighting mechanism balances the exploration of the decision boundary with the exploitation of rare classes. For example, Haocheng Luo et al. in Re-weighting Tokens: A Simple and Effective Active Learning Strategy for Named Entity Recognition apply inverse frequency weighting to individual tokens. This forces the active learning loop to prioritize rare entity tags over ubiquitous background tokens, rapidly reducing the training set's imbalance ratio.
In medical imaging, where missing a rare disease is catastrophic, Lee A. D. Cooper et al. (Focused active learning for histopathological image classification) demonstrate that weighting epistemic uncertainty by class frequency ensures the model actively seeks out rare, high-risk pathologies rather than getting distracted by common tissue variations or image artifacts.
Rather than enforcing a hard constraint on class proportions, this soft weighting—as explored by Marc Schachtsiek et al. in Class Balanced Dynamic Acquisition for Domain Adaptive Semantic Segmentation using Active Learning—allows the model to dynamically shift its budget toward underrepresented categories while still respecting high-uncertainty regions across common classes.
Pool Partitioning and Density Estimation
An alternative strategy to counteract majority-class dominance is to structurally isolate classes before sampling. If a model can generate pseudo-labels or rough predictions for the unlabeled pool, the data can be partitioned into predicted class buckets to guarantee minority representation.
Jungo Kasai demonstrates this structural approach in Low-resource Deep Entity Resolution with Transfer and Active Learning. By partitioning the unlabeled pool into predicted positive and negative classes and sampling the most uncertain examples from both partitions, the model ensures a balanced representation of difficult cases and avoids the majority-class bias inherent to entity resolution tasks.
Going a step further, Jeonghyun Lee et al. (SwiftThief: Enhancing Query Efficiency of Model Stealing by Contrastive Learning) combine uncertainty with latent feature density estimation. They apply Kernel Density Estimation in the latent space to actively pull unqueried samples from sparse regions, ensuring that minority classes are systematically uncovered even when the model's initial predictions are highly uncertain or unaligned.
Uncertainty as a Balancing Proxy Beyond Active Learning
The synergy between uncertainty and class balancing extends beyond standard active learning loops into domain adaptation and continual learning.
In unsupervised domain adaptation, Yuxi Wang et al. (Uncertainty-aware Pseudo Label Refinery for Domain Adaptive Semantic Segmentation) show that tracking target-domain uncertainty allows the system to increase the sampling frequency of source-domain images containing those high-uncertainty classes. This focuses training resources precisely where the domain gap is widest, boosting minority class accuracy.
Similarly, in long-tailed continual learning where data streams are highly skewed, Li Liu et al. (Prior-free Balanced Replay: Uncertainty-guided Reservoir Sampling for Long-Tailed Continual Learning) demonstrate that using mutual information uncertainty to select samples for a replay buffer naturally preserves minority classes. Because models remain uncertain about rare classes longer, prioritizing high-uncertainty samples automatically balances the memory buffer without requiring any prior knowledge of the global data distribution.