In the ideal world of active learning, a model identifies the most "uncertain" points, a human labels them, and the model improves rapidly. However, in the real world, data is rarely distributed evenly. When a dataset is imbalanced—where one class outnumbers another by 100 to 1—standard active learning strategies often collapse. In fact, as Umang Aggarwal, Adrian Popescu, and Céline Hudelot demonstrate in Active Learning for Imbalanced Datasets, traditional methods can actually perform worse than simple random sampling.
The research direction emerging from these papers focuses on re-engineering the acquisition function to account for class priors, ensuring that rare but critical "tail" classes are not drowned out by the majority.
The Failure of Uncertainty at the Boundary
To understand why active learning fails on imbalanced data, we must look at the decision boundary. Standard heuristics like entropy or margin sampling prioritize points where the model is most confused. In a balanced dataset, these points usually lie on the frontier between classes.
In an imbalanced scenario, the model develops a heavy bias toward the majority class. Because it sees so few minority examples, its "uncertainty" becomes poorly calibrated. As noted by Jifan Zhang and Robert D. Nowak in DIRECT: Deep Active Learning under Imbalance and Label Noise, the optimal threshold for separating classes shifts significantly. Traditional uncertainty metrics continue to sample points near a biased boundary, effectively requesting more labels for the majority class while completely missing the rare "needles in the haystack."
Balancing Informativeness with Diversity
If uncertainty alone leads to bias, the solution is to balance it with representativeness or diversity. The goal is to ensure the labeled set reflects the true underlying structure of the data, not just the model's current (and likely flawed) confusion.
Several researchers have proposed mechanisms to force this balance:
- Submodular Optimization: Ruiyu Mao and Rishabh K. Iyer introduce STONE, which uses submodular functions to select batches that are both informative and class-balanced. This is particularly vital in 3D object detection, where common objects like cars can easily overwhelm rare classes like cyclists.
- Exclusive Sparsity: Xiaoqian Wang and Heng Huang propose a New Balanced Active Learning Model that uses an exclusive sparsity norm. This mathematical constraint forces the selection process to pick samples from different underlying clusters or groups, preventing the majority class from dominating the budget.
- Deterministic Point Processes (DPPs): In Multiple-criteria Based Active Learning, Qing Li and Antoni B. Chan use $k$-DPPs to automatically balance "hardness" and "coverage," ensuring the model explores the entire feature space rather than getting stuck in high-density majority regions.
Modeling the Imbalance Explicitly
Rather than just adding a diversity "patch" to existing algorithms, some researchers argue we must model the imbalance itself.
In VaB-AL, the authors use a Variational Autoencoder (VAE) to explicitly learn the distribution of class difficulty and imbalance. By treating the class latent variables as part of a Bayesian framework, the model can identify which samples are likely to belong to rare categories before they are even labeled.
Similarly, Shyam Nuggehalli introduces DIRECT, which reduces the selection problem to a one-dimensional "optimal separation" task. This allows the algorithm to remain robust even when the data is both imbalanced and noisy, cutting annotation costs by over 60% by focusing on the samples that truly define the separation between rare and common classes.
Meta-Learning and Strategy Selection
Because no single active learning strategy works for every dataset, a new "meta" approach has emerged. Jifan Zhang and colleagues developed TAILOR, a framework that treats active learning as a multi-armed bandit problem.
TAILOR doesn't commit to one acquisition function. Instead, it tries different strategies and uses a reward function based on class diversity to decide which strategy to use for the next batch. If a strategy starts returning too many majority-class labels, the bandit shifts the budget toward a strategy that prioritizes the "tail."
Why It Works: The Intuition of "Coverage"
The core intuition across these 156 papers is a shift from refining the boundary to ensuring coverage.
In balanced data, the boundary is where the most information lives. In imbalanced data, the information is hidden in the "holes" of the feature space where the minority class might reside. By using tools like submodular optimization (STONE), triplet loss (BATL), or superpixel-based balancing (Cai et al.), these methods ensure that the active learner explores the full diversity of the dataset. This prevents the model from becoming a "specialist" in the majority class and a "failure" on the rare, often safety-critical, minority cases.