SSL (semi-supervised learning) models address class imbalance by leveraging unlabeled data and specialized techniques to prevent bias toward majority classes. Unlike supervised methods that rely solely on labeled data, SSL uses both labeled and unlabeled samples, which provides more flexibility. The core idea is to ensure the model doesn’t disproportionately favor common classes by incorporating strategies like pseudo-labeling with confidence thresholds, data augmentation, and reweighting loss functions. These methods help balance the influence of rare classes during training, even when labeled examples are scarce.
One common approach is modifying pseudo-labeling to prioritize underrepresented classes. For example, a model might generate pseudo-labels for unlabeled data but only retain predictions where the confidence exceeds a class-specific threshold. If a minority class has fewer labeled examples, the threshold could be lowered to include more of its pseudo-labels. Techniques like FixMatch use a fixed confidence threshold for all classes, but developers can adapt this by dynamically adjusting thresholds based on class frequency. Data augmentation also plays a role: applying transformations like rotation or cropping more aggressively to minority-class samples can artificially increase their effective contribution. For instance, in image classification, minority-class images might be augmented with higher variability to simulate a more balanced dataset.
Another strategy involves adjusting loss functions or incorporating class-aware weighting. SSL models often combine supervised loss (on labeled data) and unsupervised loss (on pseudo-labels). Developers can assign higher weights to minority classes in the supervised loss term or scale the unsupervised loss based on class distribution. Methods like ReMixMatch explicitly align the class distribution of pseudo-labels with the labeled data’s distribution to prevent bias. Additionally, some frameworks use consistency regularization—ensuring the model produces similar outputs for different augmentations of the same input—to reduce overfitting to majority classes. For example, a text classifier might apply synonym replacement to minority-class sentences, encouraging the model to learn robust features despite limited labeled examples. By combining these techniques, SSL models mitigate imbalance without requiring extensive labeled data. Developers can implement these ideas using libraries like PyTorch or TensorFlow, tuning parameters like augmentation strength or loss weights based on their dataset’s imbalance ratio.