Class Adaptive Thresholding in Semi Supervised Learning

478 concepts242 papers1 proposal

In semi-supervised learning (SSL) and unsupervised domain adaptation (UDA), pseudo-labeling is a remarkably intuitive strategy: let a model predict labels for unlabeled data, filter out the uncertain predictions using a confidence threshold, and retrain the model on the highly confident "pseudo-labels."

However, traditional approaches rely on a rigid, global threshold (such as requiring $\ge 95\%$ confidence) across all categories. This creates a fundamental flaw: different classes possess vastly different levels of learning difficulty. A model might quickly become highly confident about easy or high-frequency classes (like "sky" or "floor"), while remaining hesitant and less confident about rare, complex, or fine-grained classes (like specific "dog breeds" or "small objects").

As a result, a fixed global threshold creates a "rich get richer" loop. Easy classes clear the high bar effortlessly and dominate subsequent training iterations, while hard classes are systematically filtered out and ignored. This leads to severe model bias, error accumulation, and catastrophic failure on long-tailed or imbalanced datasets.

Shifting the Goalposts Dynamically

To break this loop, a powerful paradigm has emerged: class-wise adaptive thresholding. Instead of forcing every category to clear the same static bar, the training system dynamically adjusts individual thresholds based on the model's current proficiency in each class.

Fixed Global Threshold:
Class A (Easy): [====== Confidence: 96% ======] -> ACCEPTED (Threshold: 95%)
Class B (Hard): [==== Confidence: 70% ====]     -> REJECTED (Threshold: 95%)

Class-Adaptive Thresholding:
Class A (Easy): [====== Confidence: 96% ======] -> ACCEPTED (Threshold: 95%)
Class B (Hard): [==== Confidence: 70% ====]     -> ACCEPTED (Threshold: 65%)

For instance, Henry Peng Zou and Cornelia Caragea in JointMatch demonstrate that lowering thresholds for difficult classes and raising them for easier ones generates a more diverse and balanced set of pseudo-labels. By treating the threshold as an evolving parameter rather than a static constant, the model can safely explore and learn from underrepresented categories early in the training process, preventing the system from favoring easily learned categories.

Quantifying "Learning Progress"

The core engineering challenge in this paradigm is determining how to measure a model's class-specific proficiency in real time. Researchers have developed several elegant proxies to track this learning status:

1. Historical and Average Confidence

One direct approach is to monitor the model's own output distributions. Liang Yang et al. utilize the model's average confidence for each specific class as a proxy for how well it has mastered that category. If the average confidence for a class is low, the system lowers the acceptance bar, allowing the inclusion of unlabeled samples from difficult or underrepresented categories. Similarly, Yuxiang Nie et al. track the moving average of historical confidence scores to dynamically lower thresholds for challenging categories, turning model weaknesses into a mechanism for targeted data selection.

2. Relative Percentiles and Rankings

Instead of absolute confidence scores, some frameworks look at relative performance within each class. Yang Zou et al. introduced Class-Balanced Self-Training (CBST), which selects a fixed percentage of the most confident predictions within each individual category. This ensures that even if the model's top confidence for a hard class is low, the top portion of those predictions is still utilized. Teo Spadotto et al. and Yuxi Li et al. similarly leverage class-wise percentile rankings to ensure that hard-to-segment classes provide useful training signals even when their absolute confidence scores are low.

3. Internal Model Parameters and Parameters

Rather than looking only at model outputs, the selection criteria can be anchored to internal optimization states. Zhiyu Wu and Jinshi Cui take a structural approach in AllMatch, using the L2 norm of classifier weights as a proxy for class-specific learning progress. This allows the model to lower thresholds for difficult or under-represented classes based on internal optimization speeds.

4. Statistical Mixture Modeling

Instead of relying on hard heuristics, the distribution of clean versus noisy labels can be modeled statistically. Fangyuan Zhang et al. employ a Gaussian Mixture Model (GMM) to automatically learn class-specific thresholds, accounting for the varying difficulty and noise distributions inherent in different object categories during the training process.

Balancing Precision and Recall

The core intuition behind class-adaptive thresholding is a dynamic trade-off between precision and recall. A fixed high threshold prioritizes precision but destroys recall for hard classes.

To fine-tune this balance, researchers often employ non-linear scaling functions. For example, Bing Li et al. in DyConfidMatch utilize a concave mapping function to aggressively lower the confidence requirements for classes the model finds challenging, while maintaining strict, high standards for well-learned categories. This mathematical flexibility prevents minority categories from being ignored during the pseudo-labeling process while keeping noise under control.

Broad Applicability Across Modalities

This mechanism has proven vital across a wide array of complex machine learning challenges where data imbalance or domain shifts occur:

  • Fine-Grained and Complex Vision: In tasks like facial expression recognition, Jun Yu et al. show that class-adaptive thresholds prevent subtle or rare emotions from being completely excluded. For fine-grained image classification, Bowen Tian et al. use it to ensure a more balanced selection of pseudo-labels across highly similar categories.
  • 3D and Spatial Data: 3D point clouds suffer from severe sparsity and lower overall model confidence compared to 2D images. Class-level threshold adjustments allow 3D models to effectively ingest unlabeled spatial data without collapsing into majority-class predictions.
  • Decentralized and Federated Systems: As surveyed by Zixing Song et al., adaptive thresholding is crucial in federated learning to handle severe data heterogeneity and class imbalances across isolated local nodes, preventing the global model from converging on biased predictions.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 267 papers
  2. Version 1 · 2026-10-04 · 278 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_class_wise_adaptive_thresholding_6d2386a6,
  title = {class-wise adaptive thresholding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/class-wise-adaptive-thresholding/rdr_6d2386a6361247ecbb7e23d3531c5d5a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_class_wise_adaptive_thresholding_6d2386a6_v2,
  title = {class-wise adaptive thresholding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/class-wise-adaptive-thresholding/rdr_6d2386a6361247ecbb7e23d3531c5d5a/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake